Digital Heritage Archive for Memorials, Manuscripts & Ambedkar: AI-Powered Institutional Archive and Audio-Visual Knowledge Platform

1. Executive Snapshot

This dossier presents a complete technical blueprint and strategic roadmap for SIH26096, commissioned by the Ministry of Social Justice and Empowerment (MoSJE). The project addresses the historical fragmentation and accessibility barriers of Dr. B. R. Ambedkar’s writings, speeches, rare manuscripts, and Constituent Assembly debates [PS]. The proposed solution is a Centralized Hardware-Software Integrated Digital Archive Platform designed for deployment at institutions like the Dr. Ambedkar International Centre (DAIC), New Delhi [PS, KNOWN]. The platform combines interactive touch-screen kiosks, smart displays, a high-performance centralized archival server, and a specialized AI suite featuring OCR document processing, multilingual speech and translation via Bhashini APIs, vector-based semantic search, and an AI Research Assistant for constitutional Q&A [PS, INFERENCE]. This document provides a 4-6 student team with the exact architectural choices, dataset strategies, hardware configurations, and demonstration plan necessary to win Smart India Hackathon 2026.

Fundamental System Objective: To deliver an integrated hardware-software physical-digital archive that digitizes, organizes, and democratizes access to Dr. B. R. Ambedkar's intellectual legacy through interactive touch kiosks, multilingual AI audio-visual systems, and grounded RAG-based constitutional research tools [PS, INFERENCE].

2. Problem Deconstruction

The problem statement provided by MoSJE outlines nine explicit core requirements. To ensure full compliance without omitting or weakening any requirement, they are mapped below:

1. REQ-1: AI-Powered Semantic Search & Knowledge Mapping

   • Source Text: "AI-powered semantic search and intelligent knowledge mapping." [PS]
   • Interpretation: Keyword search is insufficient; the platform must understand conceptual relations (e.g., linking "Annihilation of Caste" to constitutional Article 17) using vector embeddings and knowledge graphs [INFERENCE].
   • Implementation Implication: Vector database (Qdrant / PostgreSQL with pgvector) + Graph database (Neo4j / NetworkX) for entity mapping [INFERENCE].

2. REQ-2: Full-Text & Summarized Access

   • Source Text: "Full-Text and summarized access to writings and speeches." [PS]
   • Interpretation: Users must be able to read full raw transcripts/books (e.g., BAWS 22 volumes) or generate concise, context-aware summaries for quick learning [PS, INFERENCE].
   • Implementation Implication: High-speed document reader UI + pre-computed/on-demand LLM summarization pipeline with page-accurate citation tracking [INFERENCE].

3. REQ-3: OCR Digitization of Manuscripts

   • Source Text: "OCR-based digitization of old documents and manuscripts." [PS]
   • Interpretation: Ingest low-quality, yellowed, or handwritten historical scans, remove noise, extract text, and preserve layout metadata [PS, INFERENCE].
   • Implementation Implication: Image preprocessing (OpenCV) + Indic OCR models (Tesseract 5 / PaddleOCR / Bhashini OCR API) [RESEARCHED, INFERENCE].

4. REQ-4: Multilingual Translation & Audio Narration

   • Source Text: "Multilingual translation and audio narration features." [PS]
   • Interpretation: Break language barriers across 22 scheduled Indian languages, enabling audio playback (text-to-speech) for visually impaired or non-English literate visitors [PS, RESEARCHED].
   • Implementation Implication: Integration with Digital India Bhashini NMT (Neural Machine Translation) and TTS (Text-to-Speech) APIs [RESEARCHED].

5. REQ-5: Audio-Video Archival System

   • Source Text: "Audio-video archival system for lectures, documentaries, and interviews." [PS]
   • Interpretation: Host, index, stream, and timestamp historical audio recordings, documentaries, and academic lectures with synchronized transcript playback [PS, INFERENCE].
   • Implementation Implication: HLS (HTTP Live Streaming) media server + Automated Speech Recognition (ASR) whisper/Bhashini pipeline for timestamped subtitle generation [INFERENCE].

6. REQ-6: Interactive Timeline & Memorial Storytelling

   • Source Text: "Interactive timeline and memorial story telling modules." [PS]
   • Interpretation: Visually map key milestones in Dr. Ambedkar's life (Mahad Satyagraha, Poona Pact, Drafting Committee, Conversion to Buddhism) connected to relevant documents, speeches, and photo galleries [PS, KNOWN].
   • Implementation Implication: Rich interactive UI component (React / D3.js / Vis.js) connected to chronological event schemas [INFERENCE].

7. REQ-7: AI Research Assistant

   • Source Text: "AI Research Assistant for answering queries related to Dr. Ambedkar's works and constitutional ideas." [PS]
   • Interpretation: A conversational interface that answers academic and public queries strictly based on verified historical sources without hallucinations [PS, INFERENCE].
   • Implementation Implication: Retrieval-Augmented Generation (RAG) architecture using Llama-3 / Mixtral / Claude with strict source attribution and guardrails [INFERENCE].

8. REQ-8: Hardware-Software Integrated Kiosk System

   • Source Text: "centralized hardware-software integrated digital archive system should be developed for institutions such as Dr. Ambedkar International Centre. The solution should include interactive kiosks and smart displays..." [PS]
   • Interpretation: The solution cannot be purely a web URL; it must include physical hardware kiosk integration (touchscreen interface, lockdown browser mode, edge caching, proximity sensors/accessibility options) [PS, INFERENCE].
   • Implementation Implication: Dual-mode architecture: Edge Kiosk Client (Electron / React Kiosk Shell running on Raspberry Pi 5 or Mini PC with touch display) connected to a Centralized On-Premise/Cloud Archival Server [INFERENCE].

9. REQ-9: Secure Digital Preservation & Metadata Tagging

   • Source Text: "The platform should also support secure digital preservation, metadata tagging, and institutional archival management." [PS]
   • Interpretation: Admin dashboard for archivists to upload documents, apply Dublin Core metadata standards, manage role-based access control (RBAC), and maintain archival audit logs [PS, INFERENCE].
   • Implementation Implication: Institutional admin backend with Dublin Core schema, S3/MinIO secure object storage, and checksum validation [INFERENCE].

3. Stakeholders & User Needs

4. Requirements Traceability

5. Constraints & Environment Realities

1. Hardware & Physical Kiosk Realities:

   • Kiosks deployed at high-footfall institutions like DAIC face heavy public physical interaction [PS, KNOWN]. Touches must be responsive (<200ms feedback), screens must be vandal-resistant, and the OS must run in Kiosk Lockdown Mode (preventing users from exiting to the desktop or accessing system settings) [INFERENCE].
   • Physical sound management: High ambient museum noise requires directed audio (directional speakers or dual headphone jacks) for speech listening [INFERENCE].

2. Network & Connectivity Constraints:

   • Institutional Wi-Fi/LAN can experience intermittent outages. High-resolution manuscript scans and video streaming must not crash during connectivity loss. The kiosk client must maintain a Local Edge Cache (SQLite/IndexedDB) with core offline capabilities (pre-buffered timeline, cached core documents, and UI assets) [INFERENCE].

3. Multilingual & Digital Literacy Constraints:

   • Visitors span diverse socio-economic backgrounds, age groups, and literacy levels [PS, INFERENCE]. The UI must avoid complex menu hierarchies, feature large touch targets, provide voice search (Speech-to-Text), and support primary regional scripts (Hindi, Marathi, English, Tamil, Telugu, Bengali, Gujarati) [PS, RESEARCHED].

4. Archival & Document Condition Constraints:

   • Many historical manuscripts from 1920–1950 suffer from paper yellowing, ink bleeding, torn edges, and low contrast [RESEARCHED]. Standard OCR fails on uncleaned images. Image pre-processing (binarization, deskewing, noise reduction) is mandatory prior to OCR execution [INFERENCE].

6. Research Questions

1. What exact published collections of Dr. Ambedkar are available in public digital format, and what are their copyright/licensing terms?
2. How performant are Bhashini REST/WebSocket APIs for real-time text-to-speech and translation under concurrent kiosk request spikes?
3. What is the optimal vector chunking strategy for long legal/philosophical texts (such as Constituent Assembly Debates) to maintain page-number traceability in RAG answers?
4. How can an offline kiosk client sync local telemetry/usage analytics with the central server when connectivity is restored?
5. What hardware configuration (Raspberry Pi 5 vs Intel NUC / Mini PC) offers the best cost-to-performance ratio for running local interactive touch UIs with video decoding?

7. Data Requirements & Sources

8. Existing Solutions & Gap Analysis

+---------------------------------------------------------------------------------------------------+
|                                     EXISTING SYSTEMS LANDSCAPE                                    |
+----------------------------------+----------------------------------+-----------------------------+
|    baws.in (AIC US Platform)     |    Abhilekh Patal (NAI Portal)   |    JATAN / Museums of India |
|  - Web-only text search          |  - Archival reference metadata   |  - Museum artifact catalog  |
|  - Poor scanning on old volumes  |  - No interactive AI RAG         |  - No kiosk physical specs  |
|  - No physical kiosk lockdown    |  - No voice narration / Bhashini |  - No constitutional Q&A    |
+----------------------------------+----------------------------------+-----------------------------+
                                                 |
                                                 v
+---------------------------------------------------------------------------------------------------+
|                                  PROPOSED PROTOTYPE GAP-CLOSER                                    |
|  * Integrated Touch Kiosk + Smart Display Shell (Hardware/Software)                                |
|  * Multilingual Voice AI (Bhashini STT/TTS + Indic OCR)                                           |
|  * RAG-powered Constitutional AI Assistant with exact BAWS Page Citations                          |
|  * Interactive Timeline & Concept Mapping Engine                                                  |
+---------------------------------------------------------------------------------------------------+

1. Solution 1: BAWS.in (Ambedkar International Center - AIC US Initiative)

   • Description: A web-based search portal for Babasaheb Ambedkar's writings and speeches [RESEARCHED].
   • Strengths: Allows text keyword searching across digital volumes; provides read-aloud functionality [RESEARCHED].
   • Weaknesses: Web-only layout; lacks AI semantic search, context-aware RAG assistant, physical kiosk integration, and knowledge mapping [RESEARCHED, INFERENCE].
   • Relevance & Gap: Shows demand for searchable BAWS. Proposed platform adds physical hardware kiosk capabilities, multilingual Bhashini voice pipelines, and vector-based semantic AI [INFERENCE].

2. Solution 2: Abhilekh Patal (National Archives of India)

   • Description: Official portal for access to NAI’s 100+ million archival file records [RESEARCHED].
   • Strengths: Massive database of historical catalog metadata [RESEARCHED].
   • Weaknesses: Clunky interface; no full-text search inside un-OCR'd documents; no AI summaries, voice narration, or smart kiosk deployment [RESEARCHED, INFERENCE].
   • Relevance & Gap: Abhilekh Patal is an index repository. Our solution is an immersive institutional knowledge experience [INFERENCE].

3. Solution 3: JATAN (C-DAC Pune) / Museums of India Portal

   • Description: Digital collection management software used across national museums including Dr. Babasaheb Ambedkar Museum, Pune [RESEARCHED].
   • Strengths: Strong 2D/3D museum asset cataloging and inventory control [RESEARCHED].
   • Weaknesses: Focuses on physical artifact inventory rather than deep text analysis, constitutional research, voice AI, or interactive storytelling [INFERENCE].
   • Relevance & Gap: Our system complements asset software by serving as an interactive learning and research platform [INFERENCE].

9. Candidate Solution Approaches

Approach 1: Basic Web Archive Portal (Pure Software / Keyword Search)

• Mechanism: Traditional monolithic web application (Django + PostgreSQL) with standard SQL LIKE or Full-Text Search (FTS) indexing on parsed PDF texts.
• Complexity: Low.
• Cost: Very Low.
• Evaluation:
  • Pros: Simple to build within 48 hours; minimal infrastructure cost.
  • Cons: Fails REQ-1 (no semantic concept mapping), REQ-7 (no AI assistant), REQ-8 (no kiosk hardware layer), and REQ-4 (no real-time speech/translation). Completely fails the hardware category expectation of SIH [INFERENCE].

Approach 2: Heavy Cloud-Native AI Architecture (Microservices + Commercial LLMs)

• Mechanism: Microservices on AWS/GCP using OpenAI GPT-4, Pinecone vector database, AWS Transcribe, and AWS Polly.
• Complexity: High.
• Cost: High (recurring API token costs and cloud hosting).
• Evaluation:
  • Pros: High NLP accuracy out of the box; fast rapid prototyping.
  • Cons: Severe vulnerability to internet downtime in physical museum settings; high recurring costs for public kiosks; poor performance on low-resource Indian languages compared to specialized Indic models [INFERENCE].

Approach 3: Integrated Hardware-Software Architecture with Edge Kiosk Caching & Bhashini / Indic AI Stack (RECOMMENDED)

• Mechanism: Dedicated Edge Touch-Kiosk Client (Electron/React Kiosk Shell) backed by an On-Premise/Cloud Hybrid Server (FastAPI + PostgreSQL pgvector + Qdrant). Integrates Government of India’s Bhashini API for voice/translation, Tesseract 5/PaddleOCR for document ingestion, and a localized RAG pipeline for constitutional Q&A [RESEARCHED, INFERENCE].
• Complexity: Moderate - High.
• Cost: Low (uses open-source tools and free tier government Bhashini APIs).
• Evaluation:
  • Pros: Meets 100% of problem requirements [PS]; satisfies the Hardware category requirement [PS]; robust against network drops via edge caching; zero recurring commercial LLM costs [INFERENCE].
  • Cons: Requires hardware setup and careful API pipeline integration [INFERENCE].

10. Recommended Approach & Why

Selected Approach: Approach 3 (Integrated Hardware-Software Architecture with Edge Caching & Bhashini Stack) [ASSUMPTION].

Justification for Selection:

1. Fulfills the Hardware Category Mandate: Problem SIH26096 is listed under the Hardware category [PS]. A standard web app will receive poor hardware scoring. Approach 3 incorporates physical kiosk hardware specifications, touchscreen UI design, kiosk lockdown security, and local hardware caching [INFERENCE].
2. Aligns with GoI Digital India Stack: Utilizes Bhashini APIs (Digital India Bhashini Division) for multilingual STT, NMT, and TTS, directly aligning with national technology priorities [RESEARCHED].
3. Institutional Feasibility: Places like DAIC require zero-downtime physical displays. Local edge caching guarantees that even if external WAN drops, core interactive timelines, audio streams, and pre-cached documents continue functioning smoothly [INFERENCE].

Why Alternatives Were Rejected:

• Approach 1 was rejected because simple keyword search lacks semantic understanding and completely ignores hardware/kiosk and AI requirements [INFERENCE].
• Approach 2 was rejected due to reliance on expensive commercial APIs (OpenAI/AWS), poor offline resilience, and data sovereignty concerns regarding national historical assets [INFERENCE].

11. Technical Architecture

Layered Architecture Diagram

+---------------------------------------------------------------------------------------------------+
|                                 PHYSICAL KIOSK & SMART DISPLAY LAYER                              |
|  - 32"/55" Interactive Touch Screen Display Kiosk (DAIC Visitor Gallery)                          |
|  - Kiosk Client App: Electron.js / React Touch Shell (Locked-Down Browser Environment)             |
|  - Hardware Peripherals: Thermal Printer (Quote printing), Proximity Sensor, Dual Headphone Jack |
+---------------------------------------------------------------------------------------------------+
                                                  |
                                (REST / WebSockets / Local Sync)
                                                  v
+---------------------------------------------------------------------------------------------------+
|                                    EDGE CONTROLLER / LOCAL GATEWAY                                |
|  - Raspberry Pi 5 (8GB) / Intel NUC Mini PC (Running inside Physical Kiosk)                       |
|  - Offline Edge Cache (SQLite / Local Storage for Timelines, Photos, Core Speeches)               |
|  - Kiosk Management Daemon (Heartbeat, Screen Saver, Lockdown Service)                            |
+---------------------------------------------------------------------------------------------------+
                                                  |
                                           (Secure HTTPS / WSS)
                                                  v
+---------------------------------------------------------------------------------------------------+
|                                  CENTRALIZED ARCHIVAL BACKEND SERVER                              |
|  +-----------------------------------+  +------------------------------------------------------+  |
|  |     API Gateway (FastAPI / Node)  |  |            AI / ML PROCESSING PIPELINE               |  |
|  | - Auth & Role Management (JWT)    |  | - Document OCR Engine (Tesseract 5 / OpenCV)         |  |
|  | - Archival Metadata API (Dublin C)|  | - Bhashini Pipeline (STT -> NMT -> TTS)              |  |
|  | - AV Streaming Controller (HLS)   |  | - RAG Query Engine (LangChain/LlamaIndex + Llama-3)   |  |
|  +-----------------------------------+  +------------------------------------------------------+  |
+---------------------------------------------------------------------------------------------------+
                                                  |
                                                  v
+---------------------------------------------------------------------------------------------------+
|                                      DATA STORAGE & VECTOR LAYER                                  |
|  - PostgreSQL Database (Relational Metadata, User Logs, Dublin Core Schema)                       |
|  - pgvector / Qdrant Vector Database (Document Embeddings for Semantic Search & RAG)              |
|  - MinIO / AWS S3 Object Storage (High-Res Scans, BAWS PDFs, Audio/Video Media)                   |
|  - Neo4j / NetworkX Graph DB (Knowledge Mapping of Concepts, Events, and Articles)                |
+---------------------------------------------------------------------------------------------------+

Detailed Data-Flow Narrative

1. Visitor Kiosk Session Initiated: A visitor approaches the touch kiosk at DAIC. The proximity sensor detects the visitor, waking up the screen from saver mode. The Electron touch shell presents a multilingual home screen (Hindi, English, Marathi, etc.) [INFERENCE].

2. Query / Voice Input & Processing:

   • Text Input: The user types "What were Ambedkar's views on Article 370?" into the semantic search bar.
   • Voice Input: The user presses the microphone icon and speaks in Hindi: "डॉ. आंबेडकर के शिक्षा पर क्या विचार थे?". The edge client streams audio to the backend via WebSocket. The backend routes audio to Bhashini Speech-to-Text (ASR), returning text in Hindi [RESEARCHED, INFERENCE].

3. Semantic Vector Search & RAG Generation: The text query is passed to the Embedding Model (bge-m3 / IndicBERT) to generate a dense vector [RESEARCHED]. The vector database (pgvector / Qdrant) performs similarity search across chunked BAWS volumes and Constituent Assembly Debates. The top-k relevant text chunks (with exact volume and page citations) are injected into the context window of the RAG Engine. The LLM generates a clear summary. If the visitor selected Hindi, the output is translated via Bhashini NMT and spoken aloud using Bhashini TTS [RESEARCHED, INFERENCE].

4. Archival Upload & Processing Flow (Admin Side): An archivist uploads a scanned handwritten manuscript PDF via the Admin Portal. The file is saved in MinIO Object Storage. An async Celery background task runs OpenCV binarization, executes Tesseract 5 OCR, extracts text, generates vector embeddings, creates Dublin Core metadata tags, and updates the vector index [INFERENCE].

12. Technology Stack

13. AI/ML Justification

1. Vector Embeddings & Semantic Search Engine

• Exact Purpose: Enable conceptual search (e.g., matching "social democracy" to speeches on fraternity and equality) beyond rigid keywords [PS, INFERENCE].
• Model Type: bge-m3 or IndicBERT embeddings [RESEARCHED].
• Input: Raw text user query.
• Output: Ranked array of document chunk IDs, BAWS volume numbers, and exact page references.
• Data Required: Pre-tokenized chunks of BAWS 22 volumes and Constituent Assembly Debates [PS, RESEARCHED].
• Why AI: Rule-based keyword matching misses synonyms, indirect references, and conceptual links across historical texts [INFERENCE].
• Evaluation: Mean Reciprocal Rank (MRR@10) and Normalized Discounted Cumulative Gain (NDCG) on a gold-standard query dataset [INFERENCE].
• Failure Mode & Oversight: If similarity scores fall below threshold (<0.65), fall back to exact keyword match and display: "No exact semantic match found. Showing keyword results." [INFERENCE]

2. Document OCR & Layout Processing Engine

• Exact Purpose: Convert scanned manuscript PDFs into searchable UTF-8 text while preserving document structure [PS].
• Model Type: OpenCV preprocessing + Tesseract 5 / PaddleOCR [RESEARCHED].
• Input: Raw manuscript image (PNG/TIFF/PDF).
• Output: Cleaned plain text, layout bounding boxes, and extracted metadata.
• Data Required: Historical document sample scans for OCR fine-tuning [INFERENCE].
• Why AI: Handwritten and aged printed manuscripts have non-standard fonts, uneven alignment, and paper noise that rule-based extractors cannot handle [INFERENCE].
• Evaluation: Character Error Rate (CER) and Word Error Rate (WER) against ground-truth manual transcriptions [INFERENCE].
• Failure Mode & Oversight: Flag low-confidence text outputs (CER > 15%) in the Admin Portal for manual human-in-the-loop archivist review [INFERENCE].

3. RAG AI Research Assistant ("Babasaheb AI")

• Exact Purpose: Provide natural language answers to user queries regarding Dr. Ambedkar's works and constitutional ideas [PS].
• Model Type: Quantized Llama-3-8B-Instruct / Mixtral with RAG prompt guardrails [INFERENCE].
• Input: User question + Top-5 retrieved BAWS text chunks.
• Output: Synthesized answer with inline BAWS volume/page citations.
• Data Required: Chunked vector index of official writings and speeches [PS, RESEARCHED].
• Why AI: Enables conversational exploration for non-expert visitors and synthesizes information spread across multiple volumes [INFERENCE].
• Evaluation: Answer faithfulness (Ragas framework), citation accuracy, and hallucination rate [INFERENCE].
• Failure Mode & Oversight: Strict system prompt instruction: "Answer strictly using the provided context. If the answer is not present in the historical context, state 'Information not found in available archives.'" [INFERENCE]

Where AI Should NOT Be Used:

• Historical Fact Alteration / Document Rendering: Raw historical PDFs and original manuscript scans must be served unaltered. AI must never "correct" or rephrase historical original documents [INFERENCE].
• User Authentication & Access Control: Admin permissions and kiosk lockdown security must be governed by strict deterministic RBAC, never AI decisions [INFERENCE].

14. Data Architecture & Offline/Low-Network Design

PostgreSQL Core Database Schema

-- Core Archives Table
CREATE TABLE archival_assets (
    asset_id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
    title VARCHAR(255) NOT NULL,
    asset_type VARCHAR(50) CHECK (asset_type IN ('manuscript', 'speech', 'writing', 'photo', 'audio', 'video')),
    baws_volume INT,
    page_number INT,
    author VARCHAR(100) DEFAULT 'Dr. B. R. Ambedkar',
    language VARCHAR(30) DEFAULT 'English',
    publication_date DATE,
    dublin_core_metadata JSONB NOT NULL,
    file_path VARCHAR(500) NOT NULL,
    checksum_sha256 VARCHAR(64) UNIQUE NOT NULL,
    created_at TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP
);

-- Document Chunks for Vector Search & RAG
CREATE TABLE document_chunks (
    chunk_id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
    asset_id UUID REFERENCES archival_assets(asset_id) ON DELETE CASCADE,
    chunk_index INT NOT NULL,
    raw_text TEXT NOT NULL,
    token_count INT NOT NULL,
    embedding vector(1024), -- pgvector extension column
    created_at TIMESTAMP WITH TIME ZONE DEFAULT CURRENT_TIMESTAMP
);

CREATE INDEX idx_chunks_embedding ON document_chunks USING hnsw (embedding vector_cosine_ops);

Local Caching & Offline Synchronization Flow

+-----------------------------------------------------------------------------------+
|                            LOCAL KIOSK STORAGE (SQLite / PWA)                      |
|  - Pre-loaded Static Assets: UI Shell, CSS/JS, Fonts, Icons                        |
|  - High-Priority Content: Top 50 FAQs, Panchteerth Timeline, Gallery Thumbnails   |
|  - Queue: Offline User Interaction Analytics & Kiosk Health Logs                  |
+-----------------------------------------------------------------------------------+
                                       |
                   (Network Connectivity Status Monitor)
                                       |
            +--------------------------+--------------------------+
            |                                                     |
  [ONLINE: WAN Active]                                  [OFFLINE: WAN Down]
            |                                                     |
            v                                                     v
 - Direct API calls to Central Server                  - Serve cached UI and top content
 - Bhashini Cloud TTS/STT streaming                    - Local web-speech fallback TTS
 - Dynamic RAG queries execution                       - Offline message: "AI assistant offline;
 - Sync queued offline analytics log                      exploring local cached archives."

1. Bandwidth Reduction Strategy: Images served to kiosks are automatically compressed and delivered in WebP format; audio is streamed via AAC/Ogg; video uses adaptive HLS chunks (480p/720p/1080p based on local LAN speeds) [INFERENCE].
2. Conflict Handling: Kiosks are read-only consumers. Analytical logs and feedback are queued locally with ISO timestamps and synced asynchronously using an idempotent append-only API endpoint [INFERENCE].

15. Security, Privacy & Governance

1. Personal & Visitor Data Protection: The public kiosk system operates in Anonymous Guest Mode. No personal identifiable information (PII) like name, phone number, or email is collected during visitor interactions [INFERENCE].
2. Kiosk Enclosure & OS Security: Kiosks run a locked-down Linux environment (e.g., Ubuntu Core / Kiosk OS) with disabled USB mass storage, blocked shell shortcuts (Ctrl+Alt+Del, Alt+Tab), auto-launching Electron shell, and automated nightly reboot cycles [INFERENCE].
3. Role-Based Access Control (RBAC): Admin portal access is strictly controlled via JWT authentication with multi-factor authentication (MFA) for archivists [INFERENCE]:
   • Role 1 (Viewer/Public): Read-only access to published assets.
   • Role 2 (Archivist): Upload, OCR run, edit metadata tags, draft assets.
   • Role 3 (Super Admin/DAIC Head): Publish assets, manage users, view analytics.
4. Historical Data Integrity & Provenance: Every uploaded file generates a unique SHA-256 checksum upon ingestion to detect any file tampering or corruption over time [INFERENCE].
5. AI Guardrails & Governance: RAG outputs are constrained by strict system prompts to prevent political bias, hallucinations, or off-topic conversational drift. All generated AI answers contain explicit volume and page citations for user verification [INFERENCE].

16. MVP Scope (Hackathon Build)

In Scope (Must Build for Hackathon Demo)

• Functional Kiosk Shell UI: Responsive, touch-optimized React/Electron application running in kiosk mode with high-contrast UI and multi-language support (English, Hindi, Marathi) [PS, INFERENCE].
• Digitized BAWS & Debate Corpus: Ingested and indexed 3 full volumes of BAWS (e.g., Vol 1 - Annihilation of Caste; Vol 6 - Public Finance; Vol 11 - Buddha and His Dhamma) and core Constituent Assembly Debates [PS, RESEARCHED].
• AI Vector Search & RAG Assistant ("Babasaheb AI"): Functional conversational search providing cited answers from ingested BAWS volumes [PS, INFERENCE].
• Bhashini Multilingual Speech & Narration: Integrated Bhashini REST/WebSocket API for Hindi/English voice query and speech playback [PS, RESEARCHED].
• Interactive Timeline Module: Clickable timeline covering Dr. Ambedkar’s key life milestones (Panchteerth sites) linked to relevant documents and photo galleries [PS, KNOWN].
• Manuscript OCR Processing Pipeline: Working admin page where a sample scanned manuscript PDF can be uploaded, preprocessed via OpenCV, OCR'd via Tesseract, and converted into searchable text [PS, INFERENCE].
• Archival Admin Dashboard: Basic Dublin Core metadata tagging and asset management backend [PS, INFERENCE].

Described Only / Out of Scope (Production System)

• Complete 22-volume manual catalog verification.
• Hardware integration with physical thermal receipt printers or smart room projection displays.
• Full air-gapped physical server rack installation at DAIC premises.

17. Live Demo Simulation Plan

To deliver a flawless demonstration during the SIH presentation without relying on unpredictable external venue Wi-Fi, the team will set up a local demonstration environment:

+---------------------------------------------------------------------------------------------------+
|                                   LIVE SIH DEMO ARCHITECTURE (LOCAL)                              |
+---------------------------------------------------------------------------------------------------+
|  [Hardware Kiosk Rig]                                                                             |
|  - Touchscreen Monitor / Tablet simulating 32" DAIC Kiosk Interface                               |
|  - Laptop running local FastAPI Server + PostgreSQL `pgvector` Container                            |
|  - Pre-cached Bhashini Audio responses & local fallback TTS                                       |
+---------------------------------------------------------------------------------------------------+
                                                  |
                                                  v
+---------------------------------------------------------------------------------------------------+
|                                       4-STEP DEMO STORYBOARD                                      |
|  Step 1: Visitor Storytelling Mode -> Interactive Timeline click (Mahad Satyagraha 1927)         |
|  Step 2: Voice Query -> Speaking in Hindi: "डॉ. आंबेडकर के समानता पर क्या विचार थे?"               |
|  Step 3: RAG Answer & Narration -> Real-time answer with BAWS Vol 1 Page Citation + Bhashini Speech|
|  Step 4: Archivist OCR Demo -> Live upload of yellowed historical manuscript scan -> OCR output   |
+---------------------------------------------------------------------------------------------------+

• Simulated External Dependencies: If internet latency slows Bhashini cloud calls during judging, the system smoothly falls back to pre-rendered response audio clips for demo queries while preserving the exact WebSocket data pipeline [INFERENCE].

18. Implementation Roadmap

Target Deadline: 30 September 2026 [PS]

Phase 1: Ingestion & OCR    Phase 2: RAG & Bhashini     Phase 3: Kiosk UI & Shell  Phase 4: Testing & Demo
[Oct 2025 - Dec 2025]       [Jan 2026 - Mar 2026]       [Apr 2026 - Jun 2026]      [Jul 2026 - Sep 2026]
-----------------------     ---------------------       ---------------------      ---------------------
- Scrape BAWS Vol 1-22      - Build pgvector index      - Develop React Touch UI   - Perform Kiosk load tests
- Parse PDF & Text Chunks   - Integrate Bhashini APIs   - Build Electron Shell     - User acceptance testing
- Tesseract OCR Pipeline    - Build RAG Q&A Engine      - Interactive Timeline     - SIH Demo Simulation

• Milestone 1 (31 Dec 2025): Complete corpus ingestion, PDF parsing, text chunking, and PostgreSQL database setup.
• Milestone 2 (31 Mar 2026): Functional vector search, Bhashini speech/translation integration, and RAG Q&A engine complete.
• Milestone 3 (30 Jun 2026): Touch kiosk UI frontend, Electron lockdown shell, timeline storytelling module, and admin portal operational.
• Milestone 4 (15 Sep 2026): System end-to-end integration testing, offline failover verification, and final SIH demo rehearsal.

19. Testing & Validation Plan

1. Functional Testing: Verify that all 9 problem requirements (OCR, RAG, timeline, speech, AV streaming, metadata tagging, search) pass automated unit and integration tests [INFERENCE].
2. AI Accuracy & Hallucination Testing:
   • Run 100 benchmark historical questions against the RAG assistant. Measure citation accuracy against hardcopy BAWS page numbers. Target: >92% citation precision [INFERENCE].
   • Perform adversarial testing (e.g., asking ungrounded questions like "What did Ambedkar say about smartphones?"). Ensure system outputs the guardrail response: "Information not found in archives." [INFERENCE]
3. OCR False-Positive / Character Error Testing: Test OCR extraction on degraded paper scans. Measure Character Error Rate (CER). Target: CER < 8% after OpenCV image preprocessing [INFERENCE].
4. Kiosk Usability & Performance Testing: Measure UI touch responsiveness on Raspberry Pi / Mini PC hardware. Target: Touch response < 150ms; page navigation < 300ms [INFERENCE].
5. Offline Resilience Testing: Disconnect WAN network cable during an active kiosk session. Confirm UI displays offline notice gracefully and allows exploration of pre-cached timelines and documents without crashing [INFERENCE].

20. Risk Register

21. Cost & Scalability

Hackathon Prototype Build

• Hardware: Existing team laptop + Touchscreen Display Monitor / Android Tablet as Kiosk Client (~₹0 - ₹12,000 if bought) [ASSUMPTION].
• Software & Hosting: Open-source stack (PostgreSQL, FastAPI, Docker, React, Electron); free tier Bhashini API key; local deployment (~₹0) [ASSUMPTION].
• Total Prototype Cost: ~₹0 - ₹12,000 [ASSUMPTION].

Production System (e.g., Deployment at DAIC, New Delhi)

• Kiosk Hardware: 4x Commercial 43" 4K Industrial Touch Kiosks with Intel NUC Mini PCs (~₹1,50,000 per unit x 4 = ₹6,00,000) [NEEDS VALIDATION].
• On-Premise Archival Server: 1x Dual GPU/CPU Rack Server for local vector search, media streaming, and local AI processing (~₹3,50,000) [NEEDS VALIDATION].
• Software Maintenance & Bhashini Enterprise Tier: Open-source core stack with dedicated government SLA (~₹1,00,000 / year) [NEEDS VALIDATION].
• Scalability: System architecture easily scales horizontally by deploying identical kiosk edge clients to other national memorials (e.g., Chaitya Bhoomi Mumbai, Deekshabhoomi Nagpur, Ambedkar Memorial Lucknow) connecting to the central MoSJE cloud server [INFERENCE].

22. Differentiation

23. Future Scope

1. Augmented Reality (AR) Constitution Gallery: Integrating AR markers on physical exhibits at DAIC, allowing visitors to point kiosk tablets at historical photos to play narrated video stories [INFERENCE].
2. Global Federation with International Archives: Linking DAIC archive with Columbia University and London School of Economics (LSE) archives containing Dr. Ambedkar's student records [RESEARCHED, INFERENCE].
3. Crowdsourced OCR Correction Portal: Enabling trusted researchers worldwide to review and fix OCR text transcriptions of handwritten letters [INFERENCE].

24. Assumptions & Open Items Log

Consolidated Assumptions Log

1. [ASSUMPTION] The hackathon team will use a touchscreen monitor or tablet running Electron.js to simulate the physical DAIC interactive kiosk.
2. [ASSUMPTION] Official Bhashini API keys will be accessible to the team via government developer registration for research/hackathon purposes.
3. [ASSUMPTION] Initial corpus for prototype build will be populated using publicly available BAWS PDFs from the Ministry of External Affairs portal and constitutionofindia.net.

Consolidated Open Items & Validation Log

1. [NEEDS VALIDATION] Exact hardware specs and network bandwidth available inside the physical visitor galleries at Dr. Ambedkar International Centre (DAIC), Janpath, New Delhi.
2. [NEEDS VALIDATION] Production procurement costs and vendor lead times for commercial-grade 43-inch touch kiosks in India.
3. [NEEDS VALIDATION] Formal API rate limits for Bhashini WebSocket TTS endpoints under high concurrent visitor queries.

Final Recommendation

The team should proceed with building Approach 3 (Integrated Edge-Kiosk Hardware/Software System with Bhashini AI & Vector RAG) [ASSUMPTION]. This approach delivers maximum impact, solves every explicit requirement of problem statement SIH26096, and guarantees a high score under the Hardware category by demonstrating a working touch kiosk client alongside an advanced Indic AI archival server [PS, INFERENCE].

Recommended Stack

• Kiosk Client Hardware & Shell: Raspberry Pi 5 / Mini PC + Electron.js + React.js + Tailwind CSS [INFERENCE].
• Backend API Framework: Python FastAPI [INFERENCE].
• Database Layer: PostgreSQL 16 with pgvector + MinIO Object Storage [INFERENCE].
• AI & Language Processing: Digital India Bhashini APIs (STT/NMT/TTS) + Tesseract 5 OCR + Llama-3 RAG Engine [RESEARCHED, INFERENCE].

Biggest Dependency

Access to structured UTF-8 text extractions of the 22 volumes of Babasaheb Ambedkar Writings and Speeches (BAWS) and Bhashini API keys for speech pipelines [RESEARCHED].

Biggest Risk

Bhashini API latency or failure during live judging due to venue Wi-Fi instability [INFERENCE].

Biggest Technical Challenge

Ensuring high OCR accuracy on low-contrast, historical manuscript scans without introducing character errors in regional scripts [RESEARCHED, INFERENCE].

First 5 Actions

1. Download & Parse BAWS Volumes: Download BAWS PDF Volumes 1, 6, and 11 from MEA/Archive.org and run text extraction scripts [RESEARCHED].
2. Register for Bhashini API Access: Apply for developer API credentials on bhashini.gov.in for STT, NMT, and TTS endpoints [RESEARCHED].
3. Set Up PostgreSQL + pgvector Container: Spin up a local Docker container for PostgreSQL with vector extensions enabled [INFERENCE].
4. Build Electron Touch Kiosk Shell: Scaffold a basic React + Electron kiosk client with auto-fullscreen and touch navigation [INFERENCE].
5. Implement RAG Pipeline Prototype: Build a minimal Python FastAPI endpoint using LangChain/LlamaIndex to query BAWS text chunks with volume/page citations [INFERENCE].

Validation Before Coding

• Verify that downloaded BAWS PDFs can be cleanly converted to text using PyPDF/pdfplumber [RESEARCHED].
• Test Bhashini API credentials with a simple cURL script for Hindi Speech-to-Text and Text-to-Speech execution [RESEARCHED].

Buildability Score /100

92 / 100 (Extremely high buildability for a 4-6 student team using modern Python, React, and open-source vector databases within the available timeline) [INFERENCE].

Confidence

High [PS, RESEARCHED] — All required data sources (BAWS, Constituent Assembly Debates) are public domain, and the core technology stack leverages official Digital India AI tools [PS, RESEARCHED].

Research Sources

1. Dr. B. R. Ambedkar International Centre (DAIC), Ministry of Social Justice and Empowerment

   • URL: https://daic.gov.in/ [1.1]
   • Source Type: Official Government Website
   • Supported Claim: DAIC is the premier autonomous research centre under MoSJE featuring digital library resources and archives [1.1].
   • Reliability Assessment: Highly Reliable (Official GoI Portal).

2. Dr. Babasaheb Ambedkar Writings and Speeches (BAWS) Online Center

   • URL: https://baws.in/ [2.1]
   • Source Type: Digital Archive Project
   • Supported Claim: Published BAWS volumes are available online and require keyword search, TTS, and page consistency [1.2, 2.1].
   • Reliability Assessment: Reliable Secondary Source.

3. Digital India Bhashini Division (MeitY, Government of India)

   • URL: https://bhashini.gov.in/ [4.2]
   • Source Type: Official Government AI Portal
   • Supported Claim: Bhashini provides open APIs for Speech-to-Text, Neural Machine Translation, and Text-to-Speech across 22 scheduled Indian languages [4.1, 4.4].
   • Reliability Assessment: Highly Reliable (Official GoI API Platform).

4. Constituent Assembly Debates Archive (Constitution of India / CLPR)

   • URL: https://www.constitutionofindia.net/constitution-assembly-debates/ [3.1]
   • Source Type: Academic Archive Portal
   • Supported Claim: Contains 12 volumes of 167 plenary assembly debate transcripts from 1946–1950 [3.1].
   • Reliability Assessment: Highly Reliable Academic Source.

5. National Archives of India (Abhilekh Patal Portal)

   • URL: https://www.abhilekh-patal.in/ [1.8]
   • Source Type: Official Government Archival Portal
   • Supported Claim: Houses digitized historical reference files and manuscript metadata under the Ministry of Culture [1.3, 1.8].
   • Reliability Assessment: Highly Reliable (Official GoI Portal).