Development of an AI-powered system to detect anomalies, fraud, and inefficiencies in MPLAD Scheme implementation regd.

1. Executive Snapshot

This dossier presents a complete technical, data, AI, and strategic architecture for SIH26102, commissioned by the Ministry of Statistics and Programme Implementation (MoSPI) - Data Informatics & Innovation Division (DIID) [PS]. The Members of Parliament Local Area Development Scheme (MPLADS) processes over ₹4,000 crore annually across hundreds of parliamentary constituencies, thousands of implementing agencies, and tens of thousands of individual civil works [KNOWN]. While the launch of the e-SAKSHI portal (effective April 1, 2023) and mobile application digitized the recommendation and sanction workflow [RESEARCHED], manual oversight remains overburdened by the sheer volume of unstructured cost estimates, milestone updates, geo-tagged photos, and multi-agency payment releases [INFERENCE]. This solution architecture defines an AI-powered intelligence and risk-analytics platform that ingests multi-modal data (structured financial records, text estimates, GIS locations, and photo evidence) to automatically detect duplicate works, inflated estimates, execution delays, vendor collusion, and guidelines non-compliance in real time [INFERENCE]. Designed specifically for a 4–6 student engineering team competing in Smart India Hackathon 2026, this plan bridges the gap between hackathon demonstrability and real-world government deployment [ASSUMPTION].

Fundamental System Objective:
To establish a multi-modal, real-time AI risk-scoring engine and decision-support dashboard layer for MoSPI, State Nodal Authorities, District Authorities, and MPs that automatically audits fund utilization, project progression, cost estimates, and asset creation to stop financial leakage, asset duplication, and execution delays before funds are disbursed [INFERENCE].

2. Problem Deconstruction

The official problem statement is deconstructed into explicit atomic requirements [PS] alongside derived technical implications [INFERENCE].

3. Stakeholders & User Needs

4. Requirements Traceability

5. Constraints & Environment Realities

1. Connectivity & Low-Bandwidth in Rural Areas: Implementing Agencies (IAs) and field engineers operating in remote districts often face spotty 2G/3G connectivity [KNOWN].
   Design Implication: The mobile app for field inspections must operate full offline-first with local SQLite caching, queuing photo uploads, and compressing images before sync [ASSUMPTION].
2. Fragmented & Heterogeneous Estimate Formats: Cost estimates submitted by District Engineers vary from scanned PDFs in regional languages to Excel sheets and manual text entries [KNOWN].
   Design Implication: Optical Character Recognition (OCR) pipeline with multi-lingual parsing capabilities is necessary to extract structured line-items [INFERENCE].
3. Rigid Statutory Guidelines (MPLADS 2023 Guidelines): Works must be sanctioned within 45 days of MP recommendation [RESEARCHED]. Disbursal flows via Single Nodal Account (SNA) / Central Nodal Account (CNA) on PFMS [RESEARCHED].
   Design Implication: Hardcoded rule validation layer must run BEFORE AI anomaly scoring to ensure regulatory compliance without black-box opacity [INFERENCE].
4. Data Privacy & Geolocation Sensitivity: Exact geo-coordinates of certain public/defense assets must not be leaked publicly [INFERENCE].
   Design Implication: Role-based data redaction (blurring precise coordinates on public dashboards while retaining exact coordinates for District/MoSPI audits) [ASSUMPTION].

6. Research Questions

1. What public data schemas are available on the e-SAKSHI portal (mplads.mospi.gov.in)?
   Status: Verified via web research. e-SAKSHI exposes public dashboards containing constituency-wise sanctioned works, financial release summary, and project completion status [RESEARCHED]. Raw API endpoints require validation [NEEDS VALIDATION].
2. What are the most frequent types of irregularities flagged in CAG audit reports on MPLADS?
   Status: Identified in CAG reports: (a) Sanctioning ineligible works (private properties/trusts), (b) Delay in sanctioning beyond 45 days, (c) Splitting works to avoid higher technical approvals, (d) Asset duplication across schemes (MGNREGS/State Capex), (e) Unspent funds lying with implementing agencies [RESEARCHED].
3. How are financial disbursements handled under the 2023 revised fund flow mechanism?
   Status: Disbursal operates via e-SAKSHI integrated with SBI Central Nodal Account (CNA) and PFMS, allocating drawing limits of ₹5 Crore at FY start [RESEARCHED].

7. Data Requirements & Sources

Data Categorization Summary:

• A. Explicitly Mentioned Data [PS]: Sanctions, expenditures, cost estimates, work progress, payments, asset creation.
• B. Inferred Data [INFERENCE]: Contractor/Vendor IDs, Geo-coordinates (Latitude/Longitude), Schedule of Rates (SoR) benchmark tables, SC/ST demographic percentages.
• C. Data to Collect [ASSUMPTION]: Public e-SAKSHI web dashboard records via ethical scraping.
• D. Synthetic Data [ASSUMPTION]: Simulated invoice logs with injected anomalies (e.g., duplicate work titles, inflated unit costs, impossible execution speeds).
• E. Data Needing Validation [NEEDS VALIDATION]: Direct live web-service API endpoints for internal MoSPI e-SAKSHI backend.

8. Existing Solutions & Gap Analysis

9. Candidate Solution Approaches

+---------------------------------------------------------------------------------------------------+
|                                 CANDIDATE ARCHITECTURAL APPROACHES                                 |
+-----------------------------------+----------------------------------+----------------------------+
| Approach 1: Heuristic Rule-Engine | Approach 2: Hybrid AI Engine     | Approach 3: Deep LLM Agent |
| (Deterministic Guidelines Checks) | (Rules + ML + Spatial Vector AI) | (Autonomous RAG & Vision)  |
| - High Speed, 100% Explainable    | - Balanced, Highly Accurate      | - Highly Complex, Costly   |
| - Cannot detect novel fraud       | - Recommended for SIH Hackathon  | - Unpredictable Latency    |
+-----------------------------------+----------------------------------+----------------------------+

10. Recommended Approach & Why

Selected Solution: Approach 2 — Hybrid Intelligence Engine (Deterministic Rules + Machine Learning Anomaly Detection + Geospatial/Vector Semantic Deduplication + XAI).

Why Alternatives Were Rejected:

1. Approach 1 (Pure Rule Engine) Rejected: Static thresholds cannot capture evolving fraud schemes, such as two different contractors submitting identical cost estimates under slightly reworded project titles (e.g., "Construction of Concrete Road at Ward 4" vs "PCC Paving near Ward No. IV Community Center") [INFERENCE].
2. Approach 3 (Pure LLM Agents) Rejected: Generative models are prone to hallucinating numerical values during budget auditing, suffer from high API latencies, and lack deterministic auditability required by government financial inspectors [INFERENCE].

Architectural Superiority of the Hybrid Approach:

• Layer 1 (Deterministic Rule Engine): Immediately filters hard guideline violations (e.g., 15% SC / 7.5% ST quota shortfall, projects exceeding ₹50 Lakh outside constituency, projects recommended on prohibited private lands) [RESEARCHED].
• Layer 2 (Spatial Vector & NLP Engine): Uses sentence embeddings (SBERT) + PostGIS spatial indexing to flag duplicate asset recommendations within a 200-meter radius across current and historical datasets [ASSUMPTION].
• Layer 3 (Unsupervised & Supervised ML): Uses Isolation Forest and XGBoost trained on cost-per-unit metrics (e.g., cost per meter of drainage, cost per square foot of roofing) to flag statistical cost outliers and predict completion delays [INFERENCE].
• Layer 4 (Explainable AI - XAI): Generates human-readable risk summaries using SHAP (SHapley Additive exPlanations), enabling District Collectors to understand exactly why a project scored high risk [INFERENCE].

11. Technical Architecture

Layered System Architecture Diagram

+---------------------------------------------------------------------------------------------------+
|                                     STAKEHOLDER INTERFACE LAYER                                   |
|  [MP Portal]           [District Auth Portal]      [State Nodal Portal]     [MoSPI Ministry Portal]   |
|  (React.js / PWA)      (React + Mobile App)        (Analytics Dash)         (Executive Command)      |
+---------------------------------------------------------------------------------------------------+
                                                  |
                                             REST / gRPC
                                                  v
+---------------------------------------------------------------------------------------------------+
|                                      API & GATEWAY LAYER                                          |
|  [Nginx Reverse Proxy]  --->  [FastAPI API Gateway]  --->  [RBAC & JWT Auth Guard]                  |
+---------------------------------------------------------------------------------------------------+
                                                  |
                                    Event Dispatcher (Celery)
                                                  v
+---------------------------------------------------------------------------------------------------+
|                                     HYBRID ANALYTICS CORE                                         |
|  +--------------------------+  +--------------------------+  +---------------------------------+  |
|  | Layer 1: Rule Engine     |  | Layer 2: Spatial Dedupe  |  | Layer 3: ML Anomaly & Risk      |  |
|  | - SC/ST Allocation Check |  | - SBERT Title Embeddings |  | - Isolation Forest (Cost Outlier) |  |
|  | - 45-Day Sanction SLA    |  | - PostGIS Geo-Radius Search| | - XGBoost (Delay Predictor)    |  |
|  +--------------------------+  +--------------------------+  +---------------------------------+  |
|  +----------------------------------------------------------------------------------------------+  |
|  | Layer 4: Image & Document Integrity Check (ResNet50 EXIF / OCR Rate Parser)                   |  |
|  +----------------------------------------------------------------------------------------------+  |
|  +----------------------------------------------------------------------------------------------+  |
|  | Layer 5: Explainable AI Engine (SHAP Risk Breakdown Generator)                               |  |
|  +----------------------------------------------------------------------------------------------+  |
+---------------------------------------------------------------------------------------------------+
                                                  |
                                                  v
+---------------------------------------------------------------------------------------------------+
|                                       DATA & STORAGE LAYER                                        |
|  [PostgreSQL + PostGIS]        [Vector DB: Qdrant / FAISS]       [Redis Cache & Message Queue]     |
|  (Relational Records & GIS)    (Semantic Title Vectors)          (Session & Async Queue)          |
+---------------------------------------------------------------------------------------------------+
                                                  ^
                                                  |
+---------------------------------------------------------------------------------------------------+
|                                      INGESTION & ETL PIPELINE                                     |
|  [e-SAKSHI Web Scraper]    [PFMS Payment Ledger CSV Parser]    [Estimate PDF/OCR Ingestion Engine] |
+---------------------------------------------------------------------------------------------------+

Detailed Data Flow Narrative

1. Proposal Ingestion: An MP or District Staff member inputs a work proposal (Title, Location, Category, Estimated Cost, Beneficiary Group) via the web interface or API ingestion pipeline [INFERENCE].
2. Rule Compliance Pass: The system instantly runs Layer 1 Rules. If the recommendation violates statutory guidelines (e.g., MP's cumulative SC area allocation drops below 15%), an immediate compliance warning is raised [INFERENCE].
3. Spatial & Semantic Deduplication: The proposal text is converted into a 384-dimensional dense vector using all-MiniLM-L6-v2. PostGIS performs a spatial query for existing assets within a 200-meter radius. If a high semantic similarity (>0.82) AND high spatial proximity (<200m) are found, a "Potential Duplicate Asset" alert is flagged [ASSUMPTION].
4. Cost Estimate Audit: The itemized cost estimate PDF/Excel is parsed. The system calculates cost per unit (e.g., ₹/sq ft) and runs Isolation Forest. If the unit cost deviates significantly from historical district averages (e.g., >2.5 standard deviations), it is flagged as an "Inflated Cost Estimate" [INFERENCE].
5. Predictive Delay & Risk Scoring: An XGBoost regressor evaluates historical performance of the assigned Implementing Agency, current active project load, and project magnitude to output an expected delay duration and an aggregate Project Risk Score (0–100) [INFERENCE].
6. Dashboard & Alert Dispatch: Results are written to PostgreSQL. High-risk proposals (Score > 75) trigger real-time notifications to the District Collector and MoSPI DIID dashboard for mandatory manual review before administrative sanction [INFERENCE].

12. Technology Stack

13. AI/ML Justification

+---------------------------------------------------------------------------------------------------+
|                                     AI / ML MODULE MATRIX                                         |
+--------------------------+----------------------------+-----------------------+-------------------+
| Module 1: Vector Dedupe  | Module 2: Cost Inflation   | Module 3: Delay Risk  | Module 4: CV Photo|
| (SBERT + PostGIS)        | (Isolation Forest)         | (XGBoost Classifier)  | (ResNet + EXIF)   |
| Finds duplicated titles  | Flags anomalous unit prices| Predicts project delay| Detects tampered  |
| & nearby assets          | against district norms     | & completion timelines| progress images   |
+--------------------------+----------------------------+-----------------------+-------------------+

Module 1: Semantic & Spatial Duplicate Asset Detector

1. Exact Problem Solved: Identifies duplicate work recommendations submitted across different financial years or schemes [INFERENCE].
2. Input: Proposed Project Title, Detailed Scope Text, Latitude/Longitude coordinates.
3. Output: Duplicate Probability Score (0–100%), Matched Historical Asset ID, Spatial Distance (meters).
4. Why Rule Engine Fails: Variations in titles (e.g., "Solar Street Light at Gram Panchayat 3" vs "Installation of 10W Solar Lights in GP-3") defeat simple string matching [INFERENCE].
5. Required Data: Historical list of MPLADS and state scheme assets with titles and coordinates [ASSUMPTION].
6. Evaluation Metric: Precision@K, Cosine Similarity Threshold (>0.82) [ASSUMPTION].
7. Failure Mode Handling: Low-confidence matches (50–75%) are marked as "Needs Manual Inspection" rather than auto-rejected [INFERENCE].

Module 2: Cost Estimate Inflation & Anomaly Engine

1. Exact Problem Solved: Flags estimates where material or labor unit rates exceed standard Schedule of Rates (SoR) benchmarks [INFERENCE].
2. Input: Extracted estimate line items (Item Name, Unit, Quantity, Rate per Unit).
3. Output: Anomaly Score, List of Inflated Line Items with % deviation from benchmark.
4. Why Rule Engine Fails: Price thresholds vary dynamically across districts, terrains (hilly vs plain), and inflation periods [INFERENCE].
5. Required Data: Historical SoR rate tables and past approved estimates [ASSUMPTION].
6. Evaluation Metric: Mean Absolute Error (MAE) on benchmark rates, Isolation Forest Contamination Factor = 0.05 [ASSUMPTION].

Module 3: Project Delay & Completion Time Predictor

1. Exact Problem Solved: Forecasts risk of execution delays beyond the mandated 1-year window [RESEARCHED].
2. Input: Agency Historical Performance, Sanctioned Amount, Sector, District Load, Monsoon Seasonality Index.
3. Output: Predicted Duration (Days), High-Risk Delay Alert (True/False).
4. Why Rule Engine Fails: Delay depends on complex multi-variable interactions (e.g., monsoon impact combined with high agency workload) [INFERENCE].
5. Required Data: Historical project start, sanction, and completion dates [ASSUMPTION].
6. Evaluation Metric: Receiver Operating Characteristic - Area Under Curve (ROC-AUC > 0.85) [ASSUMPTION].

Where AI Should NOT Be Used:

• Statutory Compliance Checks: Checking whether an MP has allocated 15% to SC areas and 7.5% to ST areas MUST be handled by deterministic mathematical code [RESEARCHED].
• Entitlement Balance Calculation: Calculating remaining drawing limit (₹5 Crore - Sanctioned Amount) MUST use exact ledger arithmetic, never ML [RESEARCHED].

14. Data Architecture & Offline/Low-Network Design

Core Relational Schema (PostgreSQL)

[ MP Table ] 1 --- * [ Work Proposals ] 1 --- 1 [ Sanctions ]
                           |                      |
                           | 1                    | 1
                           *                      *
                  [ Cost Estimates ]      [ Disbursements ]
                           |                      |
                           | 1                    | *
                           *                      *
                  [ Geo-Asset Images ]   [ Audit Alerts ]

Entity Specifications:

• mps: mp_id, name, house (Lok Sabha/Rajya Sabha), constituency, state, annual_entitlement_limit (50,000,000), sc_spent, st_spent.
• work_proposals: proposal_id, mp_id, work_title, sector, estimated_cost, latitude, longitude, district_id, status (Recommended/Sanctioned/Rejected/Completed), title_vector.
• cost_estimates: estimate_id, proposal_id, item_description, unit_rate, quantity, total_amount, sor_benchmark_rate, is_inflated.
• audit_alerts: alert_id, proposal_id, risk_score, alert_type (DUPLICATE_ASSET / COST_INFLATION / DELAY_RISK / GUIDELINE_VIOLATION), explanation_json, status (OPEN / RESOLVED / DISMISSED).

Offline Sync Architecture for Mobile App Field Inspections

1. Local SQLite Storage: Field inspectors record site inspection data, geo-coordinates, and capture asset photos locally on the device [ASSUMPTION].
2. Sync Queue: Outgoing requests are stored in an IndexedDB / SQLite sync_queue table with status PENDING [ASSUMPTION].
3. Chunked Upload & Compression: Photos are compressed locally to <300KB before upload [ASSUMPTION].
4. Conflict Resolution: Last-Write-Wins (LWW) based on server timestamp, with manual override option for District Engineers if server state has advanced [ASSUMPTION].

15. Security, Privacy & Governance

1. Role-Based Access Control (RBAC):
   • MP: Read/Write own recommendations; Read status of constituency works [RESEARCHED].
   • District Authority (DA): Read/Write district sanctions, review alerts, assign IAs [RESEARCHED].
   • State Nodal Authority (SNA): Read state-level aggregates, audit logs, and risk reports [RESEARCHED].
   • MoSPI Admin (DIID): Full platform access, rule threshold configuration, model retraining triggers [RESEARCHED].
2. Encryption Standards:
   • Data in Transit: TLS 1.3 for all HTTP/WebSocket traffic [KNOWN].
   • Data at Rest: AES-256 encryption for database storage and S3 photo buckets [KNOWN].
3. Audit Logging & Non-Repudiation:
   • Every administrative action (Sanction Approved, Alert Dismissed, Cost Estimate Override) is written to an append-only, immutable audit_trail_log table storing user_id, timestamp, ip_address, action, and justification_notes [INFERENCE].
4. False Positive & Decision Governance:
   • AI scores DO NOT automatically reject proposals [INFERENCE]. AI alerts serve purely as Decision Support Flags for District Collectors and Ministry Officials, preserving human oversight in administrative decisions [INFERENCE].

16. MVP Scope (Hackathon Build)

In Scope

• Multi-Role Web Dashboard: Dedicated UI views for MP, District Authority, State Nodal Authority, and Ministry Admin [ASSUMPTION].
• Core Anomaly Engines:
  1. GIS & Vector Semantic Duplicate Work Detector (PostGIS + SBERT) [ASSUMPTION].
  2. Cost Estimate Inflation Checker (Isolation Forest on itemized civil rates) [ASSUMPTION].
  3. Predictive Delay Risk Classifier (XGBoost) [ASSUMPTION].
  4. Statutory Guidelines Compliance Checker (15% SC / 7.5% ST quota + ₹5 Cr drawing limit) [ASSUMPTION].
• Explainable AI (XAI) Panel: Visual breakdown showing why a project received a high risk score [ASSUMPTION].
• Field Inspection Mobile Web / PWA: Photo capture with EXIF geo-tag validation and offline queue [ASSUMPTION].
• Interactive Geo-Spatial Map: Mapbox/Leaflet map displaying constituency assets color-coded by risk level [ASSUMPTION].

Described Only / Out of Scope

• Live production integration with internal SBI CNA banking gateway or internal PFMS production servers [NEEDS VALIDATION].
• Complete multi-lingual voice-to-text proposal entry for MPs [ASSUMPTION].
• Advanced satellite aperture radar (SAR) temporal change detection for infrastructure growth [ASSUMPTION].

17. Live Demo Simulation Plan

To present a compelling, evidence-based demonstration to SIH judges without access to live government production credentials:

1. Synthetic & Scraped Dataset Baseline: Pre-load the database with 500+ realistic MPLADS work proposals across 5 districts (combining real scraped public e-SAKSHI summary data with synthetic itemized estimates) [ASSUMPTION].
2. Live Fraud Injection Scenarios:
   • Scenario A (Duplicate Asset Injection): During the demo, create a new proposal titled "Construction of Community Hall near Bus Stand, Rampur" at coordinates (28.6139, 77.2090). The system instantly flags a Red Alert (Score: 88/100) showing a 91% match with an existing asset sanctioned 8 months prior at the exact same location [ASSUMPTION].
   • Scenario B (Inflated Cost Estimate Upload): Upload a sample estimate PDF where "Cement Concrete Road" is quoted at ₹8,500 per cubic meter (against district benchmark of ₹4,200/cu.m). The system highlights the inflated line item in red and calculates potential savings [ASSUMPTION].
   • Scenario C (Guideline Violation Trigger): Recommend a project that pushes the constituency's SC area spend down to 11% (below 15% quota). The system generates a compliance warning BEFORE submission [ASSUMPTION].
3. Role-Switching Journey: Demonstrate the workflow seamlessly transitioning across personas: MP recommends -> System flags risk -> District Collector receives alert -> Collector reviews XAI explanation and requests correction -> Ministry sees aggregate state risk score drop [ASSUMPTION].

18. Implementation Roadmap

Targeting the final deadline of 30 September 2026 [PS], the development plan spans 4 structured sprints starting from September 5, 2026 [ASSUMPTION].

+---------------------------------------------------------------------------------------------------+
|                                 IMPLEMENTATION TIMELINE & SPRINTS                                 |
+-------------------+--------------------+-----------------------+----------------------------------+
| Sprint 1: Setup   | Sprint 2: AI Core  | Sprint 3: Dashboards  | Sprint 4: Demo Prep & Validation |
| (Sept 5 - Sept 11)| (Sept 12 - Sept 18)| (Sept 19 - Sept 24)   | (Sept 25 - Sept 30)              |
| Schema, Scrapers, | ML Models, Vectors,| 4 RBAC Portals, GIS,  | End-to-End Testing, Scenarios,   |
| Rule Engine       | OCR Estimate Parser| XAI Panel Integration | Video & Pitch Deck               |
+-------------------+--------------------+-----------------------+----------------------------------+

19. Testing & Validation Plan

1. Functional Testing: Verify that all 4 RBAC roles restrict unauthorized route access; verify that rule engine flags statutory violations (15% SC, 7.5% ST, ₹5 Cr limit) with 100% accuracy [ASSUMPTION].
2. AI Anomaly & Precision Testing:
   • Duplicate Detection: Evaluate Precision & Recall on a labeled dataset of 100 pairs of duplicate/non-duplicate work descriptions. Target: Precision >= 90%, Recall >= 85% [ASSUMPTION].
   • Cost Inflation Detector: Test against synthetic estimates with known injected cost markups (+20% to +100%). Target: Detect 100% of markups > 30% [ASSUMPTION].
3. False Positive & Usability Audit: Perform threshold calibration to ensure normal variation in local material costs does not flood District Collectors with false alerts (Target: False Positive Rate < 8%) [ASSUMPTION].
4. Performance & Stress Testing: Measure API latency under simulated concurrent load (100 concurrent requests). Target: Risk scoring response time < 1.5 seconds [ASSUMPTION].
5. Offline PWA Sync Validation: Test field mobile app in airplane mode; capture photo and inspect queue; restore network and verify seamless PostgreSQL sync [ASSUMPTION].

20. Risk Register

21. Cost & Scalability

Hackathon Build

• Hosting & Cloud Compute: ₹0 (Built locally and deployed on free-tier Render / Vercel / Supabase) [ASSUMPTION].
• AI Model Inference: ₹0 (Utilizing open-source PyTorch, scikit-learn, and CPU-optimized HuggingFace embeddings running locally or on Google Colab) [ASSUMPTION].
• Total Hackathon Cost: ₹0 [ASSUMPTION].

Production Deployment (National Scale for MoSPI)

• Compute Infrastructure: Dedicated cloud instances (e.g., NIC MeghRaj Cloud or AWS India Region) with auto-scaling Kubernetes cluster [ASSUMPTION].
• Database & GIS: PostgreSQL HA cluster with PostGIS extension + managed Qdrant vector cluster [NEEDS VALIDATION].
• Storage: S3-compatible Object Storage for millions of geo-tagged inspection photos with lifecycle archival policies [NEEDS VALIDATION].
• Scalability Strategy: Asynchronous worker architecture decouples heavy OCR and vector calculations from the main API gateway, ensuring the dashboard remains responsive even during end-of-FY submission rushes [INFERENCE].

22. Differentiation

23. Future Scope

1. Automated Satellite Progress Verification: Integration with ISRO Bhuvan / Sentinel-2 satellite imagery to perform temporal vegetation and structural change detection over asset coordinates [INFERENCE].
2. Citizen Crowd-Sourced Asset Auditing: Public mobile portal allowing local citizens to rate created assets, upload ground photos, and flag maintenance issues [INFERENCE].
3. Cross-Scheme Convergence Analytics: Expanding the vector deduplication engine across state-level schemes (e.g., MLA-LADS, MGNREGS, PM-GSY) to prevent multi-funding of identical infrastructure [INFERENCE].

24. Assumptions & Open Items Log

Consolidated Assumptions [ASSUMPTION]

1. Synthetic dataset generated using realistic civil engineering line items accurately reflects actual district-level cost estimates.
2. A 200-meter geo-radius is an effective spatial threshold for detecting duplicate community asset creation in urban and semi-urban constituencies.
3. Hackathon judges evaluate solutions based on architectural depth, working multi-role UI, AI rigor, and domain alignment rather than live production government server connections.

Consolidated Open Items & Validation Needed [NEEDS VALIDATION]

1. Formal REST/gRPC API specification for e-SAKSHI backend integration in future production phases [NEEDS VALIDATION].
2. Exact cloud hosting quotas and infrastructure budget provided by NIC / MoSPI DIID for post-hackathon pilot deployment [NEEDS VALIDATION].

Final Recommendation

The team should proceed with building the proposed Hybrid AI System (Approach 2). It provides the optimal balance between high fraud detection capabilities, explainable administrative outputs, and realistic buildability for a 4–6 student team within the SIH timeframe [ASSUMPTION].

Recommended Stack

• Frontend: React.js + Tailwind CSS + Leaflet GIS (Web) & Flutter (Mobile PWA)
• Backend: FastAPI (Python 3.12) + Celery + Redis
• Databases: PostgreSQL + PostGIS (Relational + Spatial) & Qdrant / FAISS (Vector)
• AI/ML Core: Sentence-Transformers (all-MiniLM-L6-v2), scikit-learn (Isolation Forest), XGBoost, Tesseract OCR / PaddleOCR

Biggest Dependency

Availability of structured or scraped sample MPLADS project data to seed the GIS vector database and calibrate the cost anomaly engine [ASSUMPTION].

Biggest Risk

High false-positive rate in cost anomaly detection causing friction among District Engineers if local material cost variations are not properly benchmarked [ASSUMPTION].

Biggest Technical Challenge

Extracting structured itemized civil engineering rates accurately from noisy or scanned PDF estimate documents using OCR [INFERENCE].

First 5 Actions

1. Action 1: Scrape public summary tables from https://mplads.mospi.gov.in/digigov/dashboard.html to establish baseline schema and real constituency list [RESEARCHED].
2. Action 2: Initialize PostgreSQL + PostGIS database container and create core schema tables (mps, work_proposals, cost_estimates, audit_alerts).
3. Action 3: Build Python synthetic data generator script to synthesize 500+ realistic work proposals with injected duplicate titles, spatial overlaps, and inflated prices.
4. Action 4: Implement and test the SBERT + PostGIS spatial-semantic duplicate detection pipeline in FastAPI.
5. Action 5: Develop the multi-role React frontend dashboard shell with role-switching for MP, District Authority, State Nodal Authority, and MoSPI DIID.

Validation Before Coding

Verify that the all-MiniLM-L6-v2 embedding model correctly scores semantic similarity between representative civil work title variations (e.g. "Construction of PCC Road" vs "Paving of Concrete Street") above 0.80 [ASSUMPTION].

Buildability Score /100

92 / 100 (Highly feasible for a 4–6 student team using standard open-source Python and JavaScript stacks within a 3–4 week sprint window) [ASSUMPTION].

Confidence

High (Backed by official MoSPI MPLADS 2023 guidelines, verified e-SAKSHI workflow architecture, and established ML/GIS engineering patterns) [RESEARCHED].

Research Sources

1. MoSPI Official Portal - MPLADS Overview & Guidelines
   URL: https://www.mospi.gov.in/about-us/mplads
   Source Type: Official Government Source
   Supported Claim: Details revised MPLADS guidelines (2023), ₹5 Crore entitlement drawing limits, e-SAKSHI portal launch, 45-day sanction SLA, and 15% SC / 7.5% ST mandatory allocation rules.
   Reliability: Highest (Official GoI Ministry Portal)

2. Press Information Bureau (PIB) - e-SAKSHI Launch & Mobile App Releases
   URL: https://www.pib.gov.in/PressReleaseIframePage.aspx?PRID=1996730
   Source Type: Official Government Press Release
   Supported Claim: Launch of e-SAKSHI mobile application for revised fund flow procedure on 16 January 2024.
   Reliability: Highest (Official GoI Press Release)

3. e-SAKSHI Public Dashboard Portal
   URL: https://mplads.mospi.gov.in/digigov/dashboard.html
   Source Type: Official Government Data Portal
   Supported Claim: Public availability of scheme implementation metrics, constituency progress, and sanction summaries.
   Reliability: Highest (Official Scheme Portal)

4. Office of Comptroller and Auditor General of India (CAG) - Audit Reports
   URL: https://cag.gov.in/uploads/PressRelease/PR-press-brief-Report-No-22-of-2025-Eng-06944f225a58564-42853671.pdf
   Source Type: Official Audit Documentation
   Supported Claim: Documented audit findings on scheme irregularities, work completion delays, unspent balance accumulation, and scope changes.
   Reliability: Highest (Supreme Audit Institution of India)