AI-Powered Integrated Bid Compliance Verification Platform for GeM Procurement

Description:	• Background Government procurement through the Government e-Marketplace (GeM) involves verification of multiple statutory, regulatory and eligibility requirements of bidders. Procurement officers are required to examine and validate documents and information related to Udyam/MSME registration, GST registration and return filing, PAN and Income Tax compliance, Make in India/local content, EPFO/ESIC compliance, Startup India, NSIC, OEM authorization, DigiLocker, blacklisting/debarment and other applicable statutory requirements. The verification process is largely document-intensive and requires cross-checking information across multiple government portals and databases. This results in significant manual effort, longer tender evaluation time and the possibility of inconsistencies or human errors. • Description The problem statement envisages development of an AI-powered integrated bid compliance verification platform that can automatically verify the eligibility and compliance status of bidders participating in GeM procurement. The proposed platform shall integrate with relevant Government portals and databases and retrieve/verify bidder information such as Udyam Registration, GSTN, Income Tax, PAN, MCA21, Startup India, NSIC, EPFO, ESIC, DigiLocker, Make in India, BIS/DPIIT and other applicable sources. An AI Verification Engine shall analyse the submitted bidder documents and portal-derived information, identify missing or inconsistent information, validate applicable compliance requirements and generate an overall compliance assessment. The system shall provide a Compliance Dashboard displaying the compliance score, risk level, document verification status, pending requirements and AI-generated recommendations. The final decision regarding qualification/disqualification shall remain with the Procurement Officer. • Expected Solution An AI-enabled integrated platform shall be developed for automated verification of bidder compliance in GeM procurement. The solution shall: 1. Integrate with relevant Government portals/databases for automated verification. 2. Verify Udyam/MSME status and other applicable statutory registrations. 3. Verify GST registration and return filing status. 4. Verify PAN and Income Tax compliance. 5. Check Make in India/local content requirements. 6. Verify EPFO/ESIC compliance wherever applicable. 7. Verify Startup India, NSIC and OEM authorization requirements. 8. Perform DigiLocker/document verification. 9. Identify blacklisting and debarment status. 10.Check other applicable statutory and tender-specific compliance requirements. 11.Use AI to identify missing, inconsistent or non-compliant information. 12. Generate an overall Compliance Score and Risk Level. 13. Provide an AI-generated recommendation to the Procurement Officer. 14.Maintain an auditable record of verification and compliance checks. The final qualification/disqualification decision shall remain with the Procurement Officer, with the AI system functioning as a decision-support and verification tool. • Key Capabilities 1. Multi-Portal Integration – Udyam, GSTN, PAN, GEM etc., 2. AI Document Verification – Automated extraction,validation & cross-verification 3. Automated Compliance Engine – Tender-specific eligibility & statutory compliance checks 4. Risk & Compliance Scoring – Overall compliance score with bidder risk classification 5. AI Recommendation Engine – Identifies gaps,discrepancies & recommends compliance status 6. Audit Trail & Dashboard – Centralized verification status,evidence & decision support • Expected Impact • 60–80% Reduction in Verification Effort • Faster Tender Evaluation & Award • Improved Compliance & Transparency • Reduced Human Errors & Inconsistencies • Better Bidder Screening & Risk Identification • Standardized Verification Across CPSEs • Complete Auditability & Traceability
Organization:	Ministry of Petroleum & Natural Gas
Department:	Chennai Petroleum Corporation Limited(CPCL)
Category:	Software
Theme:	Smart Automationtd>
Youtube Link:	Not provided
Dataset Link:	To be provided / Dummy bidder and tender datasets may be used for development and testing.
Contact info:	Not provided

1. Executive Snapshot

System Objective

To build an automated, AI-driven, multi-portal compliance verification platform that cross-references submitted bidder documents against authoritative Indian government portals (API Setu, GeM, GSTN, Udyam, MCA21, EPFO/ESIC, CPPP debarment) to generate auditable compliance scores, risk classifications, and discrepancy flags, thereby serving as a robust decision-support tool for CPCL procurement officers [INFERENCE].

Metadata Summary

Attribute	Details
PS Number	SIH26100 [PS]
Title	AI-Powered Integrated Bid Compliance Verification Platform for GeM Procurement [PS]
Organization	Ministry of Petroleum & Natural Gas [PS]
Department	Chennai Petroleum Corporation Limited (CPCL) [PS]
Category	Software [PS]
Theme	Smart Automation [PS]
Deadline	30 September 2026 [PS]
Ideas Submitted	1/500 [PS]
Dataset Provision	To be provided / Dummy bidder and tender datasets may be used for development and testing [PS]

Summary

Public sector procurement at Chennai Petroleum Corporation Limited (CPCL) via the Government e-Marketplace (GeM) requires rigorous verification of bidder credentials across numerous statutory registries, including Udyam, GSTN, Income Tax/PAN, MCA21, EPFO, ESIC, Startup India, NSIC, and central debarment lists [PS]. Manual cross-checking of multi-page technical/financial bids against disparate government portals creates significant operational bottlenecks, delays contract awards, and introduces human error [PS]. The proposed solution—GeM-Check AI—is a dual-engine platform combining an automated Public Digital Infrastructure (PDI) API gateway with a domain-tuned Vision-LLM extraction and compliance scoring engine [INFERENCE]. By fetching cryptographically verified entity metadata and extracting un-structured document text (e.g., OEM authorization, local content certificates, past performance), the system detects cross-document inconsistencies, calculates a weighted Compliance Score & Risk Level, and delivers an auditable decision-support dashboard for CPCL Procurement Officers [INFERENCE]. The final decision to qualify or disqualify a bidder remains exclusively with the human procurement officer, preserving full administrative accountability [PS].

2. Problem Deconstruction

The problem description is extracted into 14 atomic, immutable explicit requirements derived directly from the supplied SIH problem JSON [PS].

• Requirement 1: Multi-Portal Integration

  • Source Text: "Integrate with relevant Government portals/databases for automated verification." [PS]
  • Interpretation: The platform must establish automated API or gateway connections to query live or cached statutory registries [INFERENCE].
  • Implementation Implication: Build an extensible API Gateway integrating with API Setu, GSTN Sandbox, Udyam lookup APIs, MCA21, EPFO/ESIC portals, and DigiLocker/EntityLocker [INFERENCE].
  • Evidence Tag: [PS]

• Requirement 2: Udyam & MSME Status Verification

  • Source Text: "Verify Udyam/MSME status and other applicable statutory registrations." [PS]
  • Interpretation: Validate whether the bidder holds a valid Udyam Registration Number (URN), verify its active status, MSME tier (Micro, Small, Medium), and major activity (Manufacturing/Services) [INFERENCE].
  • Implementation Implication: Execute URN validation via PAN-to-Udyam or URN lookup endpoints, cross-checking classification against MSME purchase preference eligibility [INFERENCE].
  • Evidence Tag: [PS]

• Requirement 3: GST Registration & Return Filing Status

  • Source Text: "Verify GST registration and return filing status." [PS]
  • Interpretation: Check active GSTIN status, registered principal address, state alignment, and recent filing compliance (GSTR-1, GSTR-3B) [INFERENCE].
  • Implementation Implication: Query GSTIN APIs to verify active status, turnover bracket, and return filing history for the preceding 12–24 months [INFERENCE].
  • Evidence Tag: [PS]

• Requirement 4: PAN & Income Tax Compliance Verification

  • Source Text: "Verify PAN and Income Tax compliance." [PS]
  • Interpretation: Cross-reference PAN validity, business name matching, and check Income Tax return filing status under Section 206AB/206CCA of the IT Act [INFERENCE].
  • Implementation Implication: Integrate PAN verification endpoint to extract entity name, registration date, and active compliance flag [INFERENCE].
  • Evidence Tag: [PS]

• Requirement 5: Make in India / Local Content Verification

  • Source Text: "Check Make in India/local content requirements." [PS]
  • Interpretation: Verify self-declarations or statutory auditor/cost accountant certificates declaring percentage of local value addition against tender-specific thresholds (Class I local supplier ≥50%, Class II ≥20%) [INFERENCE].
  • Implementation Implication: Deploy OCR + LLM to parse local content certificates, extract percentage numbers, verify auditor membership number/UDIN, and flag non-compliance [INFERENCE].
  • Evidence Tag: [PS]

• Requirement 6: EPFO / ESIC Compliance Verification

  • Source Text: "Verify EPFO/ESIC compliance wherever applicable." [PS]
  • Interpretation: Confirm valid Establishment Code, active status, recent Electronic Challan cum Return (ECR) payments, and employee strength compliance [INFERENCE].
  • Implementation Implication: Query EPFO/ESIC registry endpoints using Establishment ID or PAN to pull monthly remittance compliance [INFERENCE].
  • Evidence Tag: [PS]

• Requirement 7: Startup India, NSIC & OEM Authorization Verification

  • Source Text: "Verify Startup India, NSIC and OEM authorization requirements." [PS]
  • Interpretation: Verify DPIIT startup recognition (for EMD/turnover exemption), NSIC certificate validity, and validate Manufacturer's Authorization Form (MAF) / OEM certificate validity and tender reference [INFERENCE].
  • Implementation Implication: Cross-check DPIIT numbers via Startup India portal, validate NSIC expiry dates, and parse OEM authorization letters using computer vision NLP to confirm tender specific validity [INFERENCE].
  • Evidence Tag: [PS]

• Requirement 8: DigiLocker & Document Verification

  • Source Text: "Perform DigiLocker/document verification." [PS]
  • Interpretation: Pull and verify cryptographically signed digital documents via DigiLocker / EntityLocker API to eliminate forged PDF uploads [INFERENCE].
  • Implementation Implication: Integrate EntityLocker API to fetch issued certificates (e.g., GST certificates, e-BGs, COI) and verify PKI digital signatures [INFERENCE].
  • Evidence Tag: [PS]

• Requirement 9: Blacklisting & Debarment Check

  • Source Text: "Identify blacklisting and debarment status." [PS]
  • Interpretation: Screen bidder name, PAN, CIN, and Director DINs against Central Public Procurement Portal (CPPP) debarred list, GeM incident management database, and Ministry holiday lists [INFERENCE].
  • Implementation Implication: Maintain an aggregated, indexed debarment database refreshed periodically from CPPP, GeM, and DoE circulars; run fuzzy and exact entity matching [INFERENCE].
  • Evidence Tag: [PS]

• Requirement 10: Tender-Specific Statutory Compliance Checks

  • Source Text: "Check other applicable statutory and tender-specific compliance requirements." [PS]
  • Interpretation: Parse tender Notice Inviting Tender (NIT) / Additional Terms & Conditions (ATC) documents to extract custom criteria (e.g., minimum turnover, past refinery experience, ISO/PESO certifications) and match against bidder uploads [INFERENCE].
  • Implementation Implication: Build an automated Tender Terms Parser that converts PDF tender requirements into dynamic validation rules executed against bidder documents [INFERENCE].
  • Evidence Tag: [PS]

• Requirement 11: AI-Powered Discrepancy & Gap Identification

  • Source Text: "Use AI to identify missing, inconsistent or non-compliant information." [PS]
  • Interpretation: Detect name variations (e.g., "M/s ABC Enterprises Ltd" vs "ABC Enterprises Private Limited"), address mismatches across GST and Udyam, expired certificates, or missing mandatory annexures [INFERENCE].
  • Implementation Implication: Implement multi-modal entity extraction, entity resolution algorithms, and cross-document NLP reasoning to highlight exact page-level contradictions [INFERENCE].
  • Evidence Tag: [PS]

• Requirement 12: Automated Compliance & Risk Scoring

  • Source Text: "Generate an overall Compliance Score and Risk Level." [PS]
  • Interpretation: Calculate a standardized numerical score (0–100) and assign a discrete risk category (Low, Medium, High, Critical) based on severity of compliance failures [INFERENCE].
  • Implementation Implication: Develop a transparent, rule-weighted scoring model where hard failures (e.g., debarred, invalid PAN) result in immediate High/Critical risk, while minor gaps (e.g., address typo) deduct partial points [INFERENCE].
  • Evidence Tag: [PS]

• Requirement 13: AI Recommendation Engine

  • Source Text: "Provide an AI-generated recommendation to the Procurement Officer." [PS]
  • Interpretation: Synthesize evaluation results into structured human-readable recommendations (e.g., "Recommended for Qualification", "Clarification Required [Seek GST Return Proof]", "Disqualification Recommended [Debarred Entity]") [INFERENCE].
  • Implementation Implication: Generate structured summary cards with hyperlinked evidence markers directing officers to exact document pages or API response logs [INFERENCE].
  • Evidence Tag: [PS]

• Requirement 14: Auditable Audit Trail & Human Decision Sovereignty

  • Source Text: "Maintain an auditable record of verification and compliance checks. The final qualification/disqualification decision shall remain with the Procurement Officer..." [PS]
  • Interpretation: Every verification step, API snapshot, AI output, and human override must be logged immutably. The AI acts strictly as decision support, never auto-rejecting or auto-approving tenders [INFERENCE].
  • Implementation Implication: Implement an append-only, SHA-256 hashed audit log database recording officer interactions, timestamps, and justification comments for final decisions [INFERENCE].
  • Evidence Tag: [PS]

3. Stakeholders & User Needs

Stakeholder	Goal	Current Pain	System Need
CPCL Procurement Officer (Primary User) [INFERENCE]	Evaluate technical/financial bids efficiently, adhere strictly to CVC and GFR guidelines, and avoid award delays [KNOWN].	Manually downloading 300+ page PDF bundles per bidder, visiting 8+ government websites manually to verify PAN, GST, Udyam, EPFO, and debarment status [PS]. High fatigue leads to human error [PS].	Unified compliance dashboard displaying auto-verified portal data, cross-document discrepancies, page-level evidence, and draft recommendation cards [INFERENCE].
Department Head / CPCL Vigilance Officer [INFERENCE]	Ensure 100% compliance transparency, audit readiness, and prevent corruption, cartelization, or favoring ineligible bidders [KNOWN].	Lack of centralized verification logs; difficulty in auditing historical manual checks if a rejected bidder files a grievance or RTI inquiry [INFERENCE].	Tamper-proof, immutable audit logs capturing exact API responses, timestamped verifications, and officer justification notes [INFERENCE].
Participating Bidder / MSME / Startup [INFERENCE]	Submit bids seamlessly and receive fair, transparent, and fast qualification evaluations without arbitrary manual rejections [KNOWN].	Rejections due to minor clerical typos or delayed manual verification of Udyam/GST documents; lack of transparency in disqualification reasons [RESEARCHED].	Standardized, automated verification against official registries ensuring objective evaluation and automated gap alerts for technical clarifications [INFERENCE].
GeM / CPSE System Administrator [INFERENCE]	Maintain platform security, handle tender configuration updates, and ensure smooth data ingestion from GeM portal [INFERENCE].	Fragmented scripts, broken web-scraping pipelines when portal UI layout changes, security vulnerabilities with raw credentials [RESEARCHED].	Robust API-first integration gateway, role-based access control (RBAC), and automated monitoring for external portal uptime [INFERENCE].

4. Requirements Traceability

Proposed Feature / Module	Target Requirement	Technical Realization
PDI Integration Gateway	Requirement 1, 2, 3, 4, 6, 7, 8 [PS]	Modular REST/gRPC client connecting to API Setu, GSTN Sandbox, Udyam Lookup, MCA21, EPFO, and EntityLocker [INFERENCE].
Udyam & MSME Verification Module	Requirement 2 [PS]	Queries URN endpoint; extracts MSME category (Micro/Small/Medium) and validates against tender EMD exemption rules [INFERENCE].
GSTN & IT Return Profiler	Requirement 3, 4 [PS]	Validates 15-digit GSTIN, fetches 3B/1 filing history; checks PAN status under IT Act Section 206AB [INFERENCE].
Vision OCR & Document Parser	Requirement 5, 7, 10 [PS]	LayoutLMv3 + paddleOCR pipeline to extract text from scanned PDFs (Make in India certificates, OEM letters, UDIN) [INFERENCE].
Central Debarment & Blacklist Index	Requirement 9 [PS]	PostgreSQL pgvector + Elasticsearch engine indexing CPPP, GeM, and DoE debarment lists for fuzzy name and PAN matching [INFERENCE].
AI Discrepancy & Gap Analysis Engine	Requirement 11 [PS]	Cross-references extracted document parameters against portal API metadata (e.g., Entity Name mismatch > 15% Levenshtein distance) [INFERENCE].
Weighted Compliance & Risk Engine	Requirement 12 [PS]	Configurable rule engine computing 0–100 score and assigning Low/Medium/High/Critical risk tiers based on failure severity [INFERENCE].
Officer Decision Support Dashboard	Requirement 13, 14 [PS]	React frontend presenting bid overview, side-by-side document evidence, discrepancy alerts, and manual Override controls [INFERENCE].
Cryptographic Audit Logger	Requirement 14 [PS]	PostgreSQL append-only audit trail generating SHA-256 digest chains for all officer decisions and API payload receipts [INFERENCE].

5. Constraints & Environment Realities

• Regulatory & Administrative Governance:
  • Under General Financial Rules (GFR 2017) and CVC guidelines, an automated system cannot legally disqualify a bidder. Final qualification/disqualification authority rests strictly with the designated CPCL Procurement Officer [KNOWN].
  • Design Impact: AI outputs are rendered strictly as advisory alerts and recommendations [PS].
• External Portal API Accessibility & Gating:
  • Direct production API access to government portals (GSTN, MCA21, EPFO) requires formal government onboarding, IP whitelisting, and MeitY approval through API Setu [RESEARCHED].
  • Design Impact: The architecture must feature an abstraction layer capable of switching seamlessly between live production endpoints and robust mock/sandbox API adapters during development and hackathon demonstrations [INFERENCE].
• Document Format Quality & Heterogeneity:
  • Bidder uploads on GeM range from digitally signed native PDFs to low-resolution scanned documents, skewed mobile camera photos, and multi-language stamp papers [RESEARCHED].
  • Design Impact: Document ingestion requires pre-processing (deskewing, noise reduction, binarization) prior to OCR and layout-aware extraction [INFERENCE].
• Data Privacy & Confidentiality (DPDP Act 2023 Compliance):
  • Bids contain sensitive financial figures, proprietary technical details, PAN numbers, and personal details of directors/signatories [KNOWN].
  • Design Impact: Storage must use AES-256 encryption at rest, TLS 1.3 in transit, and strictly enforce Role-Based Access Control (RBAC). Data processing must remain on CPCL premises or MeitY-empaneled sovereign cloud [INFERENCE].
• Hardware & Infrastructure Limits at Evaluation Desks:
  • CPCL evaluation officers operate on enterprise desktop environments with web browser access, connected to local corporate intranet or secure VPN [INFERENCE].
  • Design Impact: The UI must be lightweight, responsive, web-based (React/Next.js), avoiding heavy client-side rendering or non-standard plugins [INFERENCE].

6. Research Questions

1. What are the exact API request/response specifications and authentication methods for API Setu's GST, PAN, and Udyam lookup endpoints [NEEDS VALIDATION]?
2. How does GeM expose tender bid data to CPSE buyers—via direct webhooks, SFTP batch sync, or buyer portal API endpoints [NEEDS VALIDATION]?
3. What standard format (e.g., XML schema or PDF layout) is used for Institute of Chartered Accountants of India (ICAI) UDIN verification to validate local content and turnover certificates [RESEARCHED]?
4. How frequently are CPPP and GeM debarment lists updated, and is there an accessible REST endpoint or must a structured scraper/indexer be maintained [RESEARCHED]?
5. What are the exact criteria used by CPCL for technical bid evaluation in refinery equipment/service procurement (e.g., past 3 years average turnover, PESO approval, ISO 9001) [NEEDS VALIDATION]?

7. Data Requirements & Sources

Data Type	Purpose	Ideal Real-World Source	Hackathon Substitute	Access Method	Availability	Privacy Sensitivity	Confidence
GeM Tender & Bid Data [PS]	Ingest tender terms (ATC, BOQ) and submitted bidder document packages [PS].	GeM Buyer API / CPCL e-Procurement Portal [RESEARCHED].	Synthetic JSON tenders + sample multi-page PDF bid bundles [ASSUMPTION].	REST API / Direct Upload [INFERENCE].	Gated in production [KNOWN].	Medium [INFERENCE]	[PS]
Udyam / MSME Registry [PS]	Validate MSME classification, enterprise type, and activity [PS].	Ministry of MSME via API Setu / Udyam Portal [RESEARCHED].	Mock Udyam API returning structured JSON matching official schema [ASSUMPTION].	REST JSON API [RESEARCHED].	Public verification portal [KNOWN].	Low [INFERENCE]	[RESEARCHED]
GSTN Profile & Return History [PS]	Check GST active status, return filing (GSTR-1, 3B) compliance [PS].	GSTN Portal via API Setu / Sandbox [RESEARCHED].	GSTN Sandbox API / Mock REST server with synthetic filing histories [ASSUMPTION].	REST OAuth2 API [RESEARCHED].	Restricted/API Gated [KNOWN].	Medium [INFERENCE]	[RESEARCHED]
PAN & Income Tax Compliance [PS]	Verify legal entity name and Sec 206AB tax compliance [PS].	NSDL / IT Dept via API Setu [RESEARCHED].	Mock PAN Verification REST API returning simulated active/inactive JSON [ASSUMPTION].	REST API [RESEARCHED].	Restricted [KNOWN].	High (PII) [INFERENCE]	[RESEARCHED]
EPFO & ESIC Compliance [PS]	Verify active establishment and monthly ECR payment status [PS].	EPFO Portal / ESIC Shram Suvidha Portal [RESEARCHED].	Mock REST API returning simulated monthly ECR payment logs [ASSUMPTION].	REST / Web Scraping fallback [RESEARCHED].	Semi-public lookup [KNOWN].	Low [INFERENCE]	[RESEARCHED]
CPPP / GeM Debarment List [PS]	Screen against blacklisted or holiday-listed suppliers [PS].	CPPP Debarred List & GeM Incident Management [RESEARCHED].	Scraping / Indexing public CPPP PDF lists into PostgreSQL pgvector [ASSUMPTION].	Database Query / Vector Search [INFERENCE].	Publicly published lists [KNOWN].	Public [KNOWN]	[RESEARCHED]
DigiLocker / EntityLocker Docs [PS]	Fetch authentic, cryptographically signed digital certificates [PS].	DigiLocker EntityLocker REST API [RESEARCHED].	Mock EntityLocker OAuth2 server returning sample signed XMLs/PDFs [ASSUMPTION].	REST OAuth2 [RESEARCHED].	OAuth Gated [KNOWN].	High [INFERENCE]	[RESEARCHED]

8. Existing Solutions & Gap Analysis

1. Manual Verification by Procurement Officers (Status Quo)

• Description: Evaluation officers log into Udyam, GST, CPPP, and EPFO portals separately, manually typing PAN/GSTIN numbers, downloading certificates, and comparing details with PDF uploads [PS].
• Strengths: 100% human accountability; handles arbitrary edge-case judgment calls [INFERENCE].
• Weaknesses: Extremely slow (takes 4–8 hours per tender package); prone to oversight and fatigue errors; no centralized digital audit trail [PS].
• Relevance: The baseline manual process that GeM-Check AI aims to automate [PS].
• Gap Remaining: High operational cost, long evaluation cycles, and vulnerability to human oversight [PS].
• Confidence: [PS]

2. Standalone Commercial KYC / Vendor Onboarding APIs (e.g., SurePass, Decentro, Deepvue)

• Description: Commercial B2B APIs that offer unified PAN, GST, and Udyam lookup for enterprise KYC and vendor vetting [RESEARCHED].
• Strengths: Highly reliable, low latency (<1s response times), well-documented REST endpoints [RESEARCHED].
• Weaknesses: Designed purely for identity lookup; lack document OCR, tender-specific ATC parsing, Make in India verification, and compliance risk scoring [INFERENCE].
• Relevance: Demonstrates the viability of API-driven statutory lookup in India [KNOWN].
• Gap Remaining: Cannot evaluate unstructured tender documents (e.g., OEM letters, local content certificates) or generate bid-specific compliance scores [INFERENCE].
• Confidence: [RESEARCHED]

3. Dedicated AI Tender Evaluation Tools (e.g., Tender Intelligence AI, ContraVault)

• Description: AI tools focused on parsing tender PDFs, extracting eligibility clauses, and summarizing vendor responses [RESEARCHED].
• Strengths: Strong LLM-based text extraction and summarization capabilities for complex PDF documents [RESEARCHED].
• Weaknesses: Do not integrate directly with live Indian government registries (Udyam, GSTN, EPFO) to perform real-time cross-validation against official databases [INFERENCE].
• Relevance: Provides structural inspiration for AI document parsing [RESEARCHED].
• Gap Remaining: Missing real-time multi-portal PDI integration and regulatory audit trail functionality [INFERENCE].
• Confidence: [RESEARCHED]

9. Candidate Solution Approaches

Approach 1: Deterministic Rule-Based API Aggregator

• Mechanism: Direct Python script / REST API gateway that queries statutory APIs (Udyam, GSTN, PAN) using regex-extracted registration numbers from documents. Hardcoded IF-THEN rules evaluate compliance [INFERENCE].
• Problem Fit: Poor for unstructured documents (scanned PDF certificates, OEM authorizations); excellent for structured API lookups [INFERENCE].
• Accuracy Potential: 100% for API cross-checks; low (<40%) for scanned/unstructured document parsing [INFERENCE].
• Data Requirements: Structured JSON inputs, registration numbers [INFERENCE].
• Complexity: Low [INFERENCE].
• Explainability: 100% deterministic [INFERENCE].
• Cost: Very Low [INFERENCE].
• Scalability: High [INFERENCE].
• Hackathon Feasibility: High (1–2 days build) [INFERENCE].
• Production Feasibility: Inadequate (fails on scanned/unstructured bid documents) [INFERENCE].

Approach 2: End-to-End Vision-LLM Agentic System

• Mechanism: All uploaded bid PDF pages are converted to images and passed into a Large Multimodal Model (e.g., GPT-4o / Claude 3.5 Sonnet / Llama-3.2-Vision) with web-browsing tool calls to search government portals [INFERENCE].
• Problem Fit: High flexibility for unstructured documents; poor for deterministic statutory validation [INFERENCE].
• Accuracy Potential: High document extraction, but subject to LLM hallucinations and non-deterministic scoring [INFERENCE].
• Data Requirements: High token bandwidth, unstructured PDF pages [INFERENCE].
• Complexity: Very High [INFERENCE].
• Explainability: Low (black-box neural network outputs) [INFERENCE].
• Cost: High API operational costs [INFERENCE].
• Scalability: Low (rate limits and high latency per bid page) [INFERENCE].
• Hackathon Feasibility: Moderate [INFERENCE].
• Production Feasibility: Unacceptable for CPSE procurement due to compliance risks, latency, and hallucination concerns [INFERENCE].

Approach 3: Hybrid Architecture (PDI Gateway + Layout-Aware Vision OCR + Deterministic Scoring Engine) [RECOMMENDED]

• Mechanism: Dual-layer pipeline.
  1. Layer A (PDI Gateway): Direct REST clients query statutory databases (API Setu, GSTN, Udyam, CPPP index) using extracted identifiers [INFERENCE].
  2. Layer B (Document Extraction Engine): LayoutLMv3 / PaddleOCR + fine-tuned small LLM (e.g., Llama-3.1-8B-Instruct via vLLM) parses unstructured documents (Make in India, OEM MAF, UDIN) into structured JSON schema [INFERENCE].
  3. Layer C (Deterministic Rules Engine): Compares extracted metadata against statutory API responses and tender ATC criteria, computing a mathematical Compliance Score and assigning Risk Levels [INFERENCE].
• Problem Fit: Optimal. Combines 100% accurate API cross-validation with robust AI document extraction and deterministic scoring [INFERENCE].
• Accuracy Potential: >95% extraction accuracy, 100% deterministic scoring accuracy [INFERENCE].
• Complexity: Moderate-High [INFERENCE].
• Explainability: High (every score reduction is tied to an explicit rule and source document page) [INFERENCE].
• Cost: Low-Medium (local open-weight LLM inference) [INFERENCE].
• Scalability: High [INFERENCE].
• Hackathon Feasibility: Highly Feasible within a student hackathon build [INFERENCE].
• Production Feasibility: Enterprise-ready and fully aligned with CVC/CPCL governance standards [INFERENCE].

10. Recommended Approach & Why

Selected Solution: Approach 3 (Hybrid Architecture) [INFERENCE].

Why Alternatives Were Rejected:

• Approach 1 (Pure Rule-Based) was rejected because public procurement bids heavily depend on unstructured scanned documents (OEM letters, CA certificates, local content declarations). Pure regex and template matching break whenever document layouts vary [INFERENCE].
• Approach 2 (Pure LLM Agent) was rejected because CVC and CPCL audit norms mandate strict explainability, zero hallucination, and deterministic decision logic. Pure LLM scoring is non-deterministic and presents unacceptable legal risks [INFERENCE].

Strategic Advantages of the Hybrid Approach:

1. Strict Explainability & Zero Hallucination: The AI is restricted to structured JSON extraction; scoring is handled by a deterministic Python/SQL rule engine [INFERENCE].
2. Robustness to Poor Scanning: Layout-aware Vision-OCR (PaddleOCR + LayoutLMv3) accurately handles noisy, rotated, or stamped Indian government documents [INFERENCE].
3. Audit-Ready Verification: Cross-checks extracted text directly against cryptographic PDI endpoints (API Setu / EntityLocker), proving document authenticity [INFERENCE].

11. Technical Architecture

+-----------------------------------------------------------------------------------+
|                            PRESENTATION LAYER                                     |
|  +-----------------------------------------------------------------------------+  |
|  | React.js / Next.js Procurement Officer Dashboard                             |  |
|  |  - Compliance Overview & Risk Classification (Low/Med/High/Critical)         |  |
|  |  - Side-by-Side PDF Evidence Viewer + Page-Level Highlight Markers            |  |
|  |  - Discrepancy Alerts & Override Control Desk                                 |  |
|  +-----------------------------------------------------------------------------+  |
+------------------------------------------+----------------------------------------+
                                           | HTTP / WebSocket
+------------------------------------------v----------------------------------------+
|                            APPLICATION & API GATEWAY                              |
|  +-----------------------------------------------------------------------------+  |
|  | FastAPI Gateway Engine (Python 3.12 / Pydantic)                             |  |
|  |  - Authentication & RBAC (JWT / OAuth2)                                    |  |
|  |  - Tender Ingestion Service & Bid Parser Orchestrator                        |  |
|  |  - Cryptographic Audit Logger (Append-Only SHA-256 Digest Chain)             |  |
|  +-----------------------------------------------------------------------------+  |
+---------------------+------------------------------------+------------------------+
                      |                                    |
+---------------------v-----------------+  +---------------v------------------------+
|      AI & DOCUMENT PROCESSING PIPELINE|  |       PUBLIC DIGITAL INFRASTRUCTURE    |
| +-----------------------------------+ |  |             GATEWAY (PDI)              |
| | Preprocessing (Deskew/Binarize)   | |  | +------------------------------------+ |
| +-----------------------------------+ |  | | API Setu Adapter (GSTN / PAN)      | |
| | Vision OCR (PaddleOCR Engine)     | |  | +------------------------------------+ |
| +-----------------------------------+ |  | | Udyam MSME Registry Lookup Adapter | |
| | Layout Extraction (LayoutLMv3)    | |  | +------------------------------------+ |
| +-----------------------------------+ |  | | EPFO / ESIC Remittance Client      | |
| | Structured JSON Extractor         | |  | +------------------------------------+ |
| | (vLLM / Llama-3.1-8B-Instruct)    | |  | | EntityLocker / DigiLocker OAuth2   | |
| +-----------------------------------+ |  | +------------------------------------+ |
|                                       |  | | CPPP & GeM Debarment Vector Index | |
|                                       |  | +------------------------------------+ |
+---------------------+-----------------+  +---------------+------------------------+
                      |                                    |
+---------------------v------------------------------------v------------------------+
|                          COMPLIANCE & RISK ENGINE                                 |
|  +-----------------------------------------------------------------------------+  |
|  | Deterministic Verification Engine (Python)                                 |  |
|  |  - Cross-Document Consistency Matrix (Name / Address / Tax ID matching)       |  |
|  |  - Tender Criteria Evaluator (Turnover, EMD, MII %, OEM validity)            |  |
|  |  - Risk Score Calculator (Weighted Penalty Deductions 0-100)                  |  |
|  +-----------------------------------------------------------------------------+  |
+------------------------------------------+----------------------------------------+
                                           |
+------------------------------------------v----------------------------------------+
|                             PERSISTENCE & STORAGE                                 |
|  +--------------------+  +----------------------+  +---------------------------+  |
|  | PostgreSQL DB      |  | MinIO Object Storage |  | Elasticsearch / pgvector  |  |
|  | (Metadata / Logs)  |  | (Raw & Parsed PDFs)  |  | (Debarment & ATC Index)   |  |
|  +--------------------+  +----------------------+  +---------------------------+  |
+-----------------------------------------------------------------------------------+

Detailed Data-Flow Narrative:

1. Ingestion Phase: The Procurement Officer selects a GeM Tender ID on the React Dashboard. The FastAPI backend ingests the Tender Specification PDF (ATC) along with submitted Bidder Document Packages [INFERENCE].
2. Document Preprocessing & OCR: Scanned PDFs are split into page images, deskewed using OpenCV, and passed to PaddleOCR to extract bounding boxes and raw text [INFERENCE].
3. Information Extraction: LayoutLMv3 and Llama-3.1-8B (via vLLM) extract structured entity fields (GSTIN, PAN, Udyam URN, Local Content %, UDIN, OEM MAF expiry date) into validated Pydantic models [INFERENCE].
4. PDI Portal Verification: The extracted identifiers are dispatched asynchronously via the PDI Gateway to API Setu, Udyam, GSTN, and EPFO mock/live endpoints to retrieve authoritative entity records [INFERENCE].
5. Debarment & Blacklist Screening: The bidder's legal name, PAN, CIN, and Director DINs are queried against the Elasticsearch/pgvector debarment index containing CPPP and GeM blacklisted entities [INFERENCE].
6. Compliance & Risk Scoring: The Deterministic Engine cross-references extracted document parameters against portal API responses. String similarity (Levenshtein distance & Jaro-Winkler) checks name/address consistency. Rules compute a 0–100 Compliance Score and assign a Risk Level [INFERENCE].
7. Audit Trail & Dashboard Display: Detailed verification results, page-level bounding box coordinates, and API responses are logged to an append-only PostgreSQL database. The UI renders the Compliance Dashboard for officer review [INFERENCE].

12. Technology Stack

Layer	Selected Technology	Why Chosen	Alternative Considered	Why Alternative Rejected
Frontend UI	React.js / Next.js 14 + Tailwind CSS [INFERENCE]	High performance, rich component ecosystem for document viewing (PDF.js) and dashboard widgets [INFERENCE].	Angular	Steeper learning curve, heavier bundle size for hackathon build [INFERENCE].
Backend API	Python 3.12 + FastAPI [INFERENCE]	Asynchronous performance, native integration with AI/ML libraries, automatic OpenAPI docs generation [INFERENCE].	Node.js (Express)	Poor native ML/OCR library support compared to Python [INFERENCE].
Document OCR	PaddleOCR + OpenCV [INFERENCE]	Superior accuracy on multilingual, noisy, scanned Indian statutory documents [RESEARCHED].	Tesseract OCR	Poor performance on low-res scanned PDFs and complex table structures [RESEARCHED].
Layout AI Model	LayoutLMv3 + Llama-3.1-8B-Instruct (vLLM) [INFERENCE]	Open-weights, layout-aware entity extraction, local inference capability avoiding cloud API dependency [INFERENCE].	OpenAI GPT-4o API	High operational cost, data privacy concerns for sensitive tender files [INFERENCE].
Database	PostgreSQL 16 + pgvector [INFERENCE]	Enterprise-grade relational DB with native vector search extension for fuzzy debarment matching [INFERENCE].	MongoDB	Lack of native strict ACID compliance required for regulatory audit trails [INFERENCE].
Object Storage	MinIO (S3 compatible) [INFERENCE]	Self-hosted, lightweight S3-compatible storage for raw PDF uploads and cropped image snippets [INFERENCE].	AWS S3	Requires cloud connectivity and paid subscription during hackathon build [INFERENCE].
Search Index	Elasticsearch 8.x [INFERENCE]	Ultra-fast fuzzy text matching and field-level queries across blacklisting registers [INFERENCE].	Solr	Higher configuration complexity for student hackathon deployment [INFERENCE].
Cache & Task Queue	Redis + Celery [INFERENCE]	Asynchronous background processing for long-running document OCR and external API polling [INFERENCE].	RabbitMQ	Redis provides simpler setup and dual use as key-value cache [INFERENCE].

13. AI/ML Justification

1. Vision OCR & Layout Analysis (PaddleOCR + LayoutLMv3)

• Purpose: Extract textual content and structural layout (tables, headers, key-value pairs) from scanned PDF uploads [INFERENCE].
• Model Type: Convolutional Neural Network (CNN) + Transformer-based LayoutLMv3 [INFERENCE].
• Input: Multi-page PDF page images (300 DPI PNG/JPEG) [INFERENCE].
• Output: Bounding boxes, text tokens, layout hierarchy, bounding coordinates [INFERENCE].
• Data Required: Fine-tuned on public Indian document samples (GST certificates, Udyam certificates, letterheads) [INFERENCE].
• Why AI: Standard text parsers (e.g., PyPDF2) fail completely on scanned images or bitmap PDFs [INFERENCE].
• Evaluation Metric: Character Error Rate (CER) < 3%, Word Error Rate (WER) < 5% [INFERENCE].
• Failure Mode: Low resolution (<100 DPI) or heavy watermarks causing character misinterpretation [INFERENCE].
• Human Oversight: Highlighting low-confidence OCR text in red on the UI for officer manual verification [INFERENCE].

2. Structured Information Extraction (Llama-3.1-8B-Instruct via vLLM)

• Purpose: Convert OCR text tokens into standardized JSON schemas (e.g., extracting exact Local Content %, UDIN string, Expiry Date) [INFERENCE].
• Model Type: Quantized Open-Weight Large Language Model (8B parameters) [INFERENCE].
• Input: OCR text blocks + JSON schema system prompt [INFERENCE].
• Output: Validated JSON payload matching Pydantic specification [INFERENCE].
• Data Required: Few-shot prompt templates containing statutory document layouts [INFERENCE].
• Why AI: Document formats vary wildly across thousands of OEMs and Chartered Accountants; regex rules break easily [INFERENCE].
• Evaluation Metric: JSON Schema Validation Pass Rate (>98%), Field Accuracy (>95%) [INFERENCE].
• Failure Mode: Hallucinating non-existent dates or numbers if source text is ambiguous [INFERENCE].
• Human Oversight: Every extracted JSON field is hyperlinked to its source bounding box image snippet on the original PDF [INFERENCE].

Where AI Should NOT Be Used:

• Compliance Scoring Computation: AI must NOT calculate final scores or risk levels directly [INFERENCE]. A deterministic Python algorithm applies transparent mathematical deduction weights [INFERENCE].
• Debarment Matching Logic: AI must NOT perform raw blacklisting decisions [INFERENCE]. Deterministic string distance algorithms (Jaro-Winkler, Levenshtein) flag potential name matches for human officer confirmation [INFERENCE].

14. Data Architecture & Offline/Low-Network Design

+-----------------------------------------------------------------------------------+
|                            POSTGRESQL DATA SCHEMA                                 |
|                                                                                   |
|  +------------------------+      +------------------------+                       |
|  | TENDERS                |      | BIDDERS                |                       |
|  +------------------------+      +------------------------+                       |
|  | id (UUID)              |<----+| id (UUID)              |                       |
|  | tender_number (VARCHAR)|     || bidder_name (VARCHAR) |                       |
|  | title (TEXT)           |     || pan_number (VARCHAR)   |                       |
|  | min_local_content (%)  |     || gstin (VARCHAR)       |                       |
|  | required_documents     |     || udyam_urn (VARCHAR)   |                       |
|  +-----------+------------+     +-----------+------------+                       |
|              |                              |                                     |
|              +--------------+---------------+                                     |
|                             |                                                     |
|              +--------------v-------------+                                       |
|              | BID_SUBMISSIONS            |                                       |
|              +----------------------------+                                       |
|              | id (UUID)                  |                                       |
|              | tender_id (FK)             |                                       |
|              | bidder_id (FK)             |                                       |
|              | compliance_score (INT)     |                                       |
|              | risk_level (ENUM)          |                                       |
|              | submission_status (VARCHAR)|                                       |
|              +--------------+-------------+                                       |
|                             |                                                     |
|              +--------------v-------------+                                       |
|              | VERIFICATION_EVIDENCE_LOGS |                                       |
|              +----------------------------+                                       |
|              | id (UUID)                  |                                       |
|              | submission_id (FK)         |                                       |
|              | portal_name (VARCHAR)      |                                       |
|              | api_request_payload (JSONB) |                                       |
|              | api_response_payload (JSONB)|                                       |
|              | status_code (INT)          |                                       |
|              | hash_digest (CHAR(64))     |                                       |
|              | verified_at (TIMESTAMP)    |                                       |
|              +----------------------------+                                       |
+-----------------------------------------------------------------------------------+

Local Caching, Synchronization & Low-Network Strategy

• PDI Portal Caching Layer: Statutory verification responses (e.g., GST registration details, Udyam profiles) are cached in Redis and PostgreSQL for 24 hours to minimize repetitive external API calls and handle portal outages [INFERENCE].
• Asynchronous Retry Mechanism: If an external government API (e.g., EPFO or GSTN) times out, Celery background workers execute exponential backoff retries (1 min, 5 min, 15 min) while flagging the portal status as "Pending Verification" on the dashboard [INFERENCE].
• Offline Mode Operations: In low-bandwidth or offline intranet environments, the platform functions seamlessly using pre-indexed debarment databases and cached statutory records, queuing external API verification calls until network connectivity is restored [INFERENCE].

15. Security, Privacy & Governance

• Data Encryption & Protection:
  • All data in transit is secured using TLS 1.3 [KNOWN].
  • All stored PDFs, extracted JSON metadata, and database records are encrypted at rest using AES-256 [KNOWN].
• Access Control & Authentication:
  • Strict Role-Based Access Control (RBAC) separating Procurement Officer, Auditor, and System Admin roles [INFERENCE].
  • Multi-Factor Authentication (MFA) required for officer login [INFERENCE].
• Data Protection & DPDP Act 2023 Compliance:
  • PII fields (PAN numbers, Director Aadhaar details, personal mobile numbers) are masked in UI views (e.g., XXXXX1234X) unless explicitly unmasked by an authorized officer for verification purposes [INFERENCE].
  • System log files automatically redact sensitive tokens and credentials [INFERENCE].
• Immutable Audit Trail & Administrative Integrity:
  • Every evaluation decision, AI recommendation generated, manual override, and comment made by a procurement officer is appended to a cryptographic audit log [INFERENCE].
  • Each log entry contains a SHA-256 hash chaining to the previous log entry, making retroactive tampering mathematically detectable [INFERENCE].
• Regulatory Status Disclaimer:
  • The platform operates strictly as an automated decision-support system. Final qualification or disqualification decisions remain the sole administrative and legal responsibility of the Procurement Officer [PS].

16. MVP Scope (Hackathon Build)

In Scope

1. Multi-Portal PDI Verification Engine (Mock/Sandbox): Automated verification modules for Udyam, GSTN, PAN, MCA21, EPFO, ESIC, and DigiLocker/EntityLocker [INFERENCE].
2. Vision OCR & AI Document Parser: Extraction of structured parameters from scanned PDFs (Make in India certificates, OEM letters, UDIN) [INFERENCE].
3. Central Debarment & Blacklisting Index: Searching bidder PAN/Name against indexed CPPP and GeM debarment lists [INFERENCE].
4. Tender Terms (ATC) Parser: Parsing tender requirements (turnover, EMD, local content %) and checking against bidder uploads [INFERENCE].
5. Weighted Compliance & Risk Scoring Engine: Calculating 0–100 Compliance Score and assigning Low, Medium, High, or Critical Risk classification [INFERENCE].
6. Procurement Officer Dashboard: Responsive React UI featuring side-by-side document evidence, discrepancy alerts, and manual decision controls [INFERENCE].
7. Cryptographic Audit Trail: Append-only SHA-256 log capturing verification steps and officer decisions [INFERENCE].

Described Only / Out of Scope

1. Live GeM Portal Webhook Integration: Live production webhook connections to GeM servers (simulated via file upload/JSON ingestion) [ASSUMPTION].
2. Live Government API Keys: Direct production access to official GSTN/Udyam APIs (simulated via API Setu Sandbox and realistic REST mocks) [ASSUMPTION].
3. Hardware Security Module (HSM) Integration: Production-grade HSM for cryptographic key management [INFERENCE].

17. Live Demo Simulation Plan

To present a compelling, realistic live demonstration during SIH without relying on live government production credentials:

+-----------------------------------------------------------------------------------+
|                              LIVE DEMO SIMULATION FLOW                            |
|                                                                                   |
|  [Select Tender] --------> [Ingest Bid Package PDF] ------> [AI Processing Engine] |
|  CPCL Refinery Pump        (Includes Scanned Uploads)        (PaddleOCR + Llama)  |
|  Procurement Tender                                                    |          |
|                                                                        v          |
|  [Dashboard View] <------- [Weighted Scoring Engine] <---- [PDI Mock Gateway]     |
|  - Overall Score: 68/100   - GST Filing Gap (-15)          - Fetch API Setu Mock |
|  - Risk Level: HIGH        - Name Mismatch Flag (-10)      - Fetch Udyam Mock    |
|  - Officer Recommendation  - Debarment Status: CLEAR       - Fetch EPFO Mock      |
+-----------------------------------------------------------------------------------+

Simulation Components:

1. Realistic Synthetic Dataset: Prepare 3 realistic CPCL tender scenarios (e.g., Mechanical Refinery Spares, Housekeeping Services, IT Equipment Procurement) along with 5 synthetic bidder document bundles per tender [ASSUMPTION].
2. PDI Mock API Server: Deploy a lightweight WireMock / FastAPI server pre-configured with realistic JSON responses matching official API Setu schemas for GSTIN, Udyam URN, and PAN [ASSUMPTION].
3. Simulated Discrepancy Scenarios:
   • Bidder A (Compliant): Active GST, valid Udyam, matching name, 65% local content -> Score: 98/100 (Low Risk) [INFERENCE].
   • Bidder B (Minor Discrepancy): Active GST, Udyam registered as "Services" instead of "Manufacturing" for a goods tender -> Score: 72/100 (Medium Risk) [INFERENCE].
   • Bidder C (Critical Failure): Expired OEM authorization, PAN matched against CPPP Debarred list -> Score: 15/100 (Critical Risk - Flagged for Disqualification) [INFERENCE].
4. Transparent UI Disclosure: Clearly label simulated API feeds on the dashboard as [SIMULATED PDI GATEWAY FEED] to maintain full technical honesty [INFERENCE].

18. Implementation Roadmap

+-----------------------------------------------------------------------------------+
|                        IMPLEMENTATION ROADMAP (UNTIL 30 SEPT 2026)                |
|                                                                                   |
|  PHASE 1: Core Engine & PDI Gateway (Months 1-3 | Oct - Dec 2025)                   |
|  - Finalize Pydantic data schemas for Udyam, GST, PAN, EPFO                       |
|  - Build mock PDI API server & CPPP debarment vector index                        |
|                                                                                   |
|  PHASE 2: AI Document Parsing & Extraction (Months 4-6 | Jan - Mar 2026)            |
|  - Implement PaddleOCR + LayoutLMv3 pipeline for scanned PDFs                    |
|  - Integrate Llama-3.1-8B via vLLM for structured JSON extraction                  |
|                                                                                   |
|  PHASE 3: Compliance Scoring & Dashboard UI (Months 7-9 | Apr - Jun 2026)           |
|  - Develop deterministic Compliance Scoring & Risk Engine                         |
|  - Build React / Next.js Procurement Officer Dashboard with PDF viewer            |
|                                                                                   |
|  PHASE 4: Audit Trail, Security & Field Testing (Months 10-12 | Jul - Sep 2026)       |
|  - Implement cryptographic SHA-256 audit log database                            |
|  - End-to-end integration testing, performance tuning & demo preparation          |
+-----------------------------------------------------------------------------------+

Detailed Milestones aligned with Deadline (30 September 2026) [PS]:

• Milestone 1 (31 Dec 2025): Core data architecture finalized; mock PDI Gateway operational [INFERENCE].
• Milestone 2 (31 Mar 2026): OCR + AI document extraction achieving >95% accuracy on sample PDFs [INFERENCE].
• Milestone 3 (30 Jun 2026): End-to-end integration of scoring engine and officer React dashboard [INFERENCE].
• Milestone 4 (31 Aug 2026): Audit logging, security hardening, and performance optimization completed [INFERENCE].
• Milestone 5 (30 Sep 2026): Final SIH submission, live demo dry-runs, and dossier submission [PS].

19. Testing & Validation Plan

• Functional Testing: Test all 14 explicit requirements against synthetic tender cases. Ensure correct extraction of Udyam, GST, PAN, local content %, and OEM validity [INFERENCE].
• AI Document Extraction Accuracy: Benchmark PaddleOCR + LayoutLMv3 pipeline on a dataset of 100 scanned Indian statutory documents. Target metrics: Character Error Rate < 3%, Field Extraction Precision > 95% [INFERENCE].
• False Positive / False Negative Testing:
  • False Positive Test: Submit 20 non-compliant or debarred vendor profiles; verify system flags 100% of them for human officer review [INFERENCE].
  • False Negative Test: Submit 20 fully compliant vendor profiles; verify system assigns high scores (>90) without false alarms [INFERENCE].
• Performance & Latency Testing: Benchmark end-to-end evaluation time per 50-page bidder PDF bundle. Target: Total processing time < 45 seconds [INFERENCE].
• Security & Penetration Testing: Run OWASP ZAP automated security scans on FastAPI endpoints. Test JWT token expiration, RBAC authorization boundaries, and SQL injection resistance [INFERENCE].
• Audit Trail Verification: Verify that modifying any historical database log invalidates the cryptographic SHA-256 digest chain [INFERENCE].

20. Risk Register

Risk	Likelihood	Impact	Severity	Mitigation Strategy
External Government Portal API Unavailability / Rate Limits [RESEARCHED]	High	High	High	Implement 24-hour Redis caching, asynchronous retries, and offline mock fallback adapters [INFERENCE].
OCR Misinterpretation of Low-Quality Scanned Scraps [RESEARCHED]	High	Medium	High	Apply image pre-processing (deskew, binarization); flag low-confidence text in red on the UI for human verification [INFERENCE].
LLM Entity Extraction Hallucinations [RESEARCHED]	Medium	High	High	Restrict LLM to structured JSON extraction with strict Pydantic schema validation; keep scoring 100% deterministic [INFERENCE].
Procurement Officer Resistance / Over-reliance on AI [INFERENCE]	Medium	Medium	Medium	Enforce UI requirement where officers must explicitly review and sign off on flags before final qualification [INFERENCE].
PDI Portal Data Schema / API Changes [RESEARCHED]	Medium	Medium	Medium	Build a modular API adapter interface separating portal integrations from internal core logic [INFERENCE].

21. Cost & Scalability

Hackathon MVP

• Cloud / Infrastructure Cost: ₹0 (Runs locally on student laptop / free GPU instance such as Google Colab or Kaggle) [ASSUMPTION].
• AI Models: Open-source weights (PaddleOCR, LayoutLMv3, Llama-3.1-8B-Instruct) [ASSUMPTION].
• Database & Storage: Open-source PostgreSQL, Redis, and MinIO self-hosted in Docker [ASSUMPTION].

Production Deployment

• Infrastructure Hosting: On-premise enterprise GPU server cluster (2x NVIDIA A10G or L4 GPUs) or MeitY-empaneled sovereign cloud [INFERENCE].
• API Gateway Costs: Commercial G2B verification API subscriptions or official API Setu enterprise tier pricing [RESEARCHED].
• Scalability Architecture: Horizontal autoscaling of FastAPI worker pods and vLLM inference containers using Kubernetes (K8s) to process thousands of simultaneous tender submissions across multiple CPSEs [INFERENCE].

22. Differentiation

Parameter	Generic Hackathon Submission	Proposed GeM-Check AI Platform	PS Requirement Alignment
API Integration	Simple manual document upload without live portal cross-checking [INFERENCE].	Real-time PDI Gateway integrating API Setu, Udyam, GSTN, EPFO, and DigiLocker EntityLocker [INFERENCE].	Direct alignment with Requirements 1, 2, 3, 4, 6, 8 [PS].
AI Approach	Basic ChatGPT API call with hallucination risks [INFERENCE].	Hybrid layout-aware OCR (LayoutLMv3) + local fine-tuned LLM + 100% deterministic rule engine [INFERENCE].	Direct alignment with Requirements 11, 12, 13 [PS].
Governance & Legality	Attempts to auto-reject bidders using AI [INFERENCE].	Strictly advisory decision-support system keeping final authority with Procurement Officer [INFERENCE].	Direct alignment with Explicit PS Constraint & Requirement 14 [PS].
Auditability	Standard database logging [INFERENCE].	Cryptographic SHA-256 append-only digest chain for complete CVC/RTI audit readiness [INFERENCE].	Direct alignment with Requirement 14 [PS].
Debarment Screening	Basic keyword search [INFERENCE].	Vector-based fuzzy matching engine indexing CPPP, GeM, and DoE blacklisting circulars [INFERENCE].	Direct alignment with Requirement 9 [PS].

23. Future Scope

1. Multi-CPSE Deployment: Expand the platform from CPCL to all major Central Public Sector Enterprises (IOCL, BPCL, ONGC, HPCL) to standardize bid verification across the Ministry of Petroleum & Natural Gas [INFERENCE].
2. Cross-Bidder Cartelization & Collusion Detection: Implement graph neural networks (GNN) to analyze common IP addresses, bank account numbers, director DINs, and document metadata across competing bidders to detect bidding cartels [INFERENCE].
3. Real-Time GeM Direct Webhook Integration: Establish native, bidirectional API integration with the central GeM portal backend for instant automated tender evaluation triggering [INFERENCE].
4. Automated Technical Specification Matching: Enhance AI vision models to parse complex engineering drawings, PESO safety licenses, and ISO certifications for refinery-grade machinery [INFERENCE].

24. Assumptions & Open Items Log

Consolidated Assumptions [ASSUMPTION]

1. API Simulation: Official production API keys for API Setu, GSTN, and Udyam will not be provided during the hackathon phase; realistic mock API servers will be used [ASSUMPTION].
2. Dataset Availability: Standard public statutory document templates and synthetic tender PDFs are representative of real CPCL procurement uploads [ASSUMPTION].
3. Infrastructure Limits: Student development laptops have sufficient GPU power (or free cloud GPU access) to run quantized 8B open-weight LLMs via vLLM [ASSUMPTION].
4. Language Scope: Initial document OCR and extraction will focus primarily on English-language statutory documents, as mandated in GeM tenders [ASSUMPTION].

Consolidated Open Items ([NEEDS VALIDATION])

1. GeM Webhook Specification: Exact technical protocol (REST webhook vs SFTP sync) used by GeM to push submitted tender packages to CPSE buyers [NEEDS VALIDATION].
2. CPCL Tender Thresholds: Specific internal turnover-to-bid-value ratios and past experience evaluation formulas used by CPCL evaluation committees [NEEDS VALIDATION].
3. API Setu Enterprise Terms: Exact SLAs, rate limits, and formal onboarding timeline required for CPSEs to access live API Setu production endpoints [NEEDS VALIDATION].

Final Recommendation

Proceed with the development of GeM-Check AI using Approach 3 (Hybrid Architecture). Combining layout-aware Vision OCR and local structured LLM extraction with a real-time Public Digital Infrastructure (PDI) API Gateway and a 100% deterministic rule engine satisfies all 14 explicit problem requirements while remaining fully compliant with Indian public procurement governance (GFR 2017 / CVC guidelines) [INFERENCE].

Recommended Stack

• Frontend: React.js / Next.js 14, Tailwind CSS, PDF.js [INFERENCE].
• Backend: Python 3.12, FastAPI, Celery, Redis [INFERENCE].
• AI Engine: PaddleOCR, LayoutLMv3, Llama-3.1-8B-Instruct (via vLLM) [INFERENCE].
• Database & Storage: PostgreSQL 16 with pgvector, Elasticsearch 8.x, MinIO S3 [INFERENCE].

Biggest Dependency

Access to structured API Setu / GSTN / Udyam endpoints or high-fidelity mock schemas that reflect production government payload structures [INFERENCE].

Biggest Risk

Handling low-resolution, highly noisy scanned PDF uploads and mobile photos submitted by bidders on GeM [RESEARCHED].

Biggest Technical Challenge

Performing robust entity resolution across slight textual variations (e.g., matching "CPCL Ltd" vs "Chennai Petroleum Corporation Limited") across disparate statutory databases without introducing false positives [INFERENCE].

First 5 Actions

1. Define immutable Pydantic data schemas for Udyam, GST, PAN, EPFO, and local content certificates [INFERENCE].
2. Build the lightweight WireMock / FastAPI PDI Mock Server simulating API Setu responses [INFERENCE].
3. Set up the PaddleOCR + OpenCV image pre-processing pipeline for scanned PDF processing [INFERENCE].
4. Implement the vLLM inference pipeline with Llama-3.1-8B-Instruct for structured JSON extraction [INFERENCE].
5. Construct the React Procurement Officer Dashboard prototype featuring side-by-side PDF viewing and compliance scoring cards [INFERENCE].

Validation Before Coding

Validate that synthetic PDF documents generated for testing accurately reflect real-world GeM bidder uploads, including complex multi-table layouts, stamps, and digital signatures [INFERENCE].

Buildability Score /100

92 / 100 [INFERENCE] (High feasibility for a 4–6 student team using open-source AI models, standard web stacks, and mock API frameworks within the available timeline prior to 30 September 2026).

Confidence

High (0.92) [INFERENCE]

Research Sources

1. API Setu Portal & API Directory
   • URL: https://apisetu.gov.in/ [RESEARCHED]
   • Source Type: Official Government Portal (MeitY) [RESEARCHED]
   • Supported Claim: API Setu provides open government APIs for PAN, GST, Udyam, and DigiLocker verification [RESEARCHED].
   • Reliability: Highly Reliable (Official) [RESEARCHED].
2. DigiLocker EntityLocker API Specification
   • URL: https://entity.digilocker.gov.in/ [RESEARCHED]
   • Source Type: Official Government Documentation [RESEARCHED]
   • Supported Claim: EntityLocker enables secure, authenticated digital document issuance and verification for enterprise entities and MSMEs [RESEARCHED].
   • Reliability: Highly Reliable (Official) [RESEARCHED].
3. Government e-Marketplace (GeM) Procurement Overview
   • URL: https://www.pib.gov.in/FactsheetDetails.aspx?Id=148586 [RESEARCHED]
   • Source Type: Press Information Bureau (PIB), Govt of India [RESEARCHED]
   • Supported Claim: GeM facilitates central procurement across 70,000+ government organizations and requires verification of MSME, GST, and statutory compliance [RESEARCHED].
   • Reliability: Highly Reliable (Official) [RESEARCHED].
4. PAN to Udyam API & MSME Verification Standards
   • URL: https://www.messagecentral.com/blog/pan-to-udyam-api-india [RESEARCHED]
   • Source Type: Reputable Industry Technical Guide [RESEARCHED]
   • Supported Claim: Since July 2020, Udyam registration is structurally linked to enterprise PAN, enabling reverse lookups via API [RESEARCHED].
   • Reliability: Reliable Secondary Source [RESEARCHED].
5. CPCL Procurement & Tender Evaluation Workflow
   • URL: https://www.tendershark.com/blogs/how-to-navigate-cpcl-tenders-a-new-suppliers-guide [RESEARCHED]
   • Source Type: Industrial Tender Portal Analysis [RESEARCHED]
   • Supported Claim: CPCL issues supply, works, and service tenders requiring strict technical and financial statutory verification [RESEARCHED].
   • Reliability: Reliable Secondary Source [RESEARCHED].