Skip to content

PRD_AI_Scheme_Matching.md

Latest commit

 

History

History
284 lines (226 loc) · 14.8 KB

PRD_AI_Scheme_Matching.md

File metadata and controls

284 lines (226 loc) · 14.8 KB
Raw

Product Requirements Document

AI-Driven Scheme Matching for Marginalized Entrepreneurs

SIH Problem Statement: AI-Driven Scheme Matching for Marginalized Entrepreneurs Document Version: 1.0 Status: Draft for Hackathon Build

1. Overview

1.1 Problem Statement

Marginalized entrepreneurs — women, SC/ST/OBC individuals, persons with disabilities, minorities, and rural/informal-sector founders — face a fragmented, jargon-heavy landscape of hundreds of government schemes (central, state, and sector-specific) offering grants, subsidies, and loans. Most eligible entrepreneurs never discover the schemes they qualify for, either due to information overload, language barriers, or low digital literacy.

1.2 Product Vision

A single platform where an entrepreneur (or an intermediary acting on their behalf) enters basic profile details and receives a ranked, plain-language list of government schemes they are actually eligible for — with a clear "what you get, what you need, how to apply" breakdown, available in their own language.

1.3 Objectives

  • Reduce time-to-discovery of relevant schemes from hours of manual search to under 2 minutes
  • Make eligibility criteria understandable without requiring legal/bureaucratic literacy
  • Serve users directly or through intermediaries (NGOs, Common Service Centres, bank correspondents)
  • Demonstrate genuine AI/NLP-driven matching, not a static filtered list

2. Target Users / Personas

Persona Description Primary Need
Direct User — Rural Entrepreneur Low-to-moderate digital literacy, may prefer voice/regional language Simple guided intake, spoken results
Direct User — Urban/Semi-Urban Entrepreneur Comfortable with apps, wants speed and accuracy Fast, precise matching with minimal input
Intermediary — NGO/CSC Worker Processes many applicants on behalf of communities Bulk profile handling, case tracking dashboard
Intermediary — Bank Correspondent Assists with loan-linked scheme applications Document checklist, application status tracking

3. Scope

3.1 In Scope (Hackathon MVP)

  • Guided profile intake form (category, location, sector, business stage)
  • Rule-based eligibility filtering engine
  • NLP-based semantic matching (business description → scheme description)
  • Plain-language scheme summarization
  • Ranked results with document checklists
  • Basic regional-language text + voice output (1–2 languages demoed)
  • Curated static dataset of 15–25 real schemes

3.2 Out of Scope (Post-Hackathon / Future Roadmap)

  • Live scraping/auto-sync with government scheme portals
  • Full 22-language coverage
  • User authentication and account persistence
  • Direct in-app application submission (redirect to official portals instead)
  • Full NGO/CSC caseworker dashboard with multi-user roles
  • Payment or subsidy disbursement tracking

4. Features & Functional Requirements

4.1 Smart Profile Intake

  • FR-1: User can input: category (General/SC/ST/OBC/Minority/Women/PwD), state/district, business sector, business stage (idea/early/existing), approximate funding need
  • FR-2: User can optionally provide a free-text business description ("I make handwoven textiles and want to sell online")
  • FR-3: Voice input option using speech-to-text for users who prefer not to type
  • FR-4: Intake form must be completable in under 5 steps / 2 minutes

4.2 AI-Powered Matching Engine

  • FR-5: System applies hard-constraint rule filtering first (category, state, income ceiling, sector match) to narrow the scheme pool
  • FR-6: System applies semantic similarity matching (sentence embeddings) between the user's free-text business description and each scheme's description/purpose text, to catch non-keyword matches
  • FR-7: Combined score (rule-match confidence + semantic similarity) produces a ranked list, highest relevance first
  • FR-8: Each result displays a relevance indicator (e.g., "Strong Match" / "Possible Match")

4.3 Scheme Explainer

  • FR-9: Each matched scheme is summarized into three fields: What you get, What you need, How to apply — max 2–3 lines each
  • FR-10: Summaries are pre-generated (batch, offline) for the curated dataset to ensure demo reliability; live LLM summarization is a stretch goal only

4.4 Multilingual & Voice Output

  • FR-11: Scheme summaries available in at least one regional language in addition to English/Hindi for the demo
  • FR-12: Text-to-speech playback of scheme summaries

4.5 Document Checklist

  • FR-13: Each matched scheme displays a checklist of commonly required documents (Aadhaar, caste certificate, business registration proof, bank passbook, etc.)
  • FR-14: Link to the scheme's official application portal

4.6 Intermediary View (Stretch Goal)

  • FR-15: A simplified caseworker screen to enter multiple applicant profiles and view their matches in a list, for NGO/CSC use

4.7 Optional Stretch: Document Verification (CNN)

  • FR-16 (stretch, optional): If time permits, an OCR/CNN-based module to auto-read uploaded ID/certificate images and pre-fill category or eligibility fields. This is a nice-to-have that demonstrates additional technical depth — not required for the core value proposition, and should only be attempted after FR-1 through FR-14 are solid.

5. Non-Functional Requirements

Category Requirement
Performance Matching results returned in under 3 seconds for a dataset of ~25 schemes
Usability Interface usable by someone with minimal digital literacy — large touch targets, minimal text, icon-supported navigation
Reliability (for demo) Core matching flow must work offline / without live API dependency wherever possible, to avoid demo failure from network issues
Accessibility Voice output and simple language support are core accessibility features, not add-ons
Scalability (design intent, not built for MVP) Architecture should be describable as scalable to hundreds of schemes and multiple languages, even if the prototype only demonstrates a subset

6. System Architecture

┌──────────────────────────────┐
│     Frontend (React +        │
│     Tailwind CSS)            │
│  - Profile intake form       │
│  - Results / scheme cards    │
│  - Voice input/output widget │
└───────────────┬──────────────┘
                │ REST API (JSON)
┌───────────────▼──────────────┐
│   Backend API (FastAPI /     │
│   Node.js + Express)         │
│  - Auth-free session handling│
│  - Orchestrates matching     │
└───────┬───────────────┬──────┘
        │               │
┌───────▼──────┐ ┌──────▼────────────┐
│ Rule Engine   │ │ NLP Matching Layer │
│ (Python logic │ │ (Sentence-BERT /   │
│  / decision   │ │  Sentence-         │
│  tables)      │ │  Transformers      │
│               │ │  embeddings +      │
│               │ │  cosine similarity)│
└───────┬──────┘ └──────┬────────────┘
        │               │
┌───────▼───────────────▼──────┐
│   Scheme Database              │
│   (PostgreSQL or MongoDB)      │
│  - Scheme metadata             │
│  - Eligibility rules            │
│  - Pre-generated summaries     │
└────────────────────────────────┘

Supporting services:
- Translation (Bhashini API / Google Translate API) — regional language text
- Text-to-Speech (gTTS or Bhashini TTS) — voice output
- (Stretch) OCR/CNN document reader — Tesseract OCR + lightweight CNN classifier

6.1 Tech Stack Summary

Layer Technology Rationale
Frontend React + Tailwind CSS Fast to build, component-driven, matches modern UI expectations
UI Design Reference Dribbble (mood board / inspiration only) + a component library (e.g., shadcn/ui) for actual build Dribbble is inspiration, not a code library — pair it with a real component kit
Backend FastAPI (Python) or Node/Express Python preferred if the NLP layer is in Python, for simpler integration
Database PostgreSQL (structured scheme/eligibility data) Rule-based filtering benefits from relational queries
NLP Matching Sentence-Transformers (e.g., all-MiniLM-L6-v2) Free, fast, runs locally without heavy GPU needs — ideal for hackathon
Summarization Pre-generated via LLM API (batch, offline) Avoids live-call latency/failure risk during the demo
Translation/TTS Bhashini API (Govt. of India's language AI initiative) or Google Translate/TTS Directly relevant given the "marginalized/regional" framing — using an Indian government-backed language API is also a strong narrative point for judges
Optional OCR Tesseract OCR + simple CNN classifier Only if time permits; supports FR-16
Project Management Notion Team task tracking, PRD hosting, sprint board — not part of shipped product
Not Used GAN No generative-image or synthetic-data need in this problem; omitted to avoid unjustified complexity

7. Data Model (Simplified)

Scheme

scheme_id, name, issuing_body, sector, states_applicable[],
eligible_categories[], min_age, max_age, income_ceiling,
business_stage_applicable[], benefit_description,
eligibility_text_raw, plain_summary_what, plain_summary_need,
plain_summary_apply, required_documents[], official_link,
regional_summaries: { hi: "...", <lang_code>: "..." }

UserProfile (session-based, not persisted long-term for MVP)

category, state, district, sector, business_stage,
funding_need_range, business_description_text

MatchResult

scheme_id, rule_match_score, semantic_similarity_score,
combined_relevance_score, match_label ("Strong"/"Possible")

8. Core User Flow

  1. User lands on the app → selects "Start" (or "Speak to Start" for voice mode)
  2. Guided intake: category → state → sector → business stage → (optional) free-text description
  3. System runs rule filter → narrows scheme pool
  4. System runs semantic matching on remaining pool using the free-text description
  5. Results screen: ranked scheme cards, each showing relevance label, plain-language summary, document checklist, official link
  6. User can toggle language / play audio summary per scheme
  7. (Stretch) Caseworker mode: repeat steps 2–6 for multiple applicants, saved in a session list

9. API Design (Core Endpoints)

Method Endpoint Purpose
POST /api/profile Submit user profile, returns session/profile ID
GET /api/schemes/match?profile_id=... Returns ranked matched schemes
GET /api/schemes/{scheme_id} Full detail for one scheme
GET /api/schemes/{scheme_id}/summary?lang=... Localized plain-language summary
POST /api/schemes/{scheme_id}/tts?lang=... Returns audio stream of the summary
POST /api/caseworker/applicants (Stretch) Add applicant under caseworker session

10. Success Metrics (for pitch/demo narrative)

Metric Target (illustrative, for pitch)
Time to receive matched results < 2 minutes from landing to results
Relevant schemes surfaced per user 3–5 high-confidence matches on average
Language accessibility At least 2 languages demoed (English + 1 regional)
Reduction in irrelevant scheme noise From "100s of schemes to search" → "top 3–5 ranked"

11. Risks & Assumptions

Risk Mitigation
Live LLM/translation API failure during demo Pre-generate and cache all summaries/translations used in the demo path
"AI" claim seen as superficial by judges Make the semantic-matching step visibly demonstrable — show a business description matching a scheme that doesn't share exact keywords
Scheme data becomes outdated Acknowledge in pitch as a "sync pipeline" roadmap item; not built in MVP
Overscoping (trying to build caseworker dashboard + OCR + everything) Explicitly sequence build priority: core matching flow first, stretch goals only if time remains

12. Build Timeline (36-Hour Hackathon Sequencing)

Phase Hours Deliverable
Setup + dataset curation 0–4 Repo scaffolding, 15–25 real schemes structured in DB
Rule engine + intake form 4–10 Working filter on hard-constraint fields
NLP semantic matching 10–16 Embedding-based ranking layered on top of rule filter
Summaries + document checklist 16–20 Pre-generated plain-language content wired to results
Frontend polish (Tailwind + UI) 20–28 Clean, accessible results screen; mobile-responsive
Multilingual + voice layer 28–32 1–2 languages, TTS working end-to-end
Testing + demo script rehearsal 32–36 Full persona walkthrough tested twice, backup screenshots ready

13. Team Roles (Suggested for a 6-Member SIH Team)

Role Responsibility
Backend/API Lead FastAPI/Node endpoints, rule engine
ML/NLP Lead Embedding matching, similarity scoring
Frontend Lead React + Tailwind UI
Data/Content Lead Scheme dataset curation, plain-language summaries
Language/Accessibility Lead Translation + TTS integration
PM/Presentation Lead Notion tracking, pitch deck, demo script

14. Appendix: Sample Scheme Entry (Reference Format)

{
  "scheme_id": "stand_up_india",
  "name": "Stand-Up India Scheme",
  "issuing_body": "Department of Financial Services, Govt. of India",
  "sector": "General/Manufacturing/Services/Trading",
  "states_applicable": ["All India"],
  "eligible_categories": ["SC", "ST", "Women"],
  "min_age": 18,
  "business_stage_applicable": ["idea", "early"],
  "benefit_description": "Bank loans between ₹10 lakh and ₹1 crore for greenfield enterprises",
  "plain_summary_what": "A bank loan of ₹10 lakh–₹1 crore to start a new business.",
  "plain_summary_need": "You must be SC/ST or a woman, aged 18+, starting a new (not existing) business.",
  "plain_summary_apply": "Apply through any scheduled commercial bank branch or the Stand-Up India portal.",
  "required_documents": ["Aadhaar", "Caste certificate (if applicable)", "Business plan", "Address proof"],
  "official_link": "https://www.standupmitra.in"
}

This PRD is scoped for hackathon MVP delivery. Sections marked "stretch" or "future roadmap" should only be pursued once the core matching flow (FR-1 through FR-14) is fully working and demo-ready.