🛡️ Regulatory Intel

AI compliance radar for cross-border remittance.

Tracks every circular from RBI · IFSCA · MCA · SEBI · FATF — auto-classifies relevance, extracts action items, and lets compliance staff interrogate any document in plain English.

Live Demo · Features · Quick Start · API · Architecture

---

Recruiter Quick Scan

Signal	Details
Product	AI regulatory monitoring desk for remittance/compliance teams
What it demonstrates	Full-stack delivery, resilient scraping, document pipelines, AI fallback design, operator-grade workflow thinking
Differentiator	Evidence-cited AI analysis rather than black-box summaries
Stack	Next.js, React, TypeScript, SQLite, Gemini 2.5 Flash, Groq, PDF parsing, OCR

---

🎯 What is this?

A production-grade regulatory monitoring system built for Glomopay, a GIFT City IFSC-licensed outward-remittance operator.

Every day, RBI, IFSCA, MCA, SEBI, and FATF publish circulars. Most are noise. A handful change how a remittance company has to operate — tomorrow. Regulatory Intel finds the signal, scores its relevance, and tells your compliance team exactly what to do next.

Built as a take-home assignment for the Glomopay Full Stack Builder role. Now public as a reference architecture for any fintech that needs auditable AI-assisted compliance tooling.

---

🧩 Product Snapshot

Layer	What ships	Portfolio signal
Regulatory ingestion	RBI RSS, IFSCA API, MCA, SEBI, and FATF fetchers with per-source health reporting.	Shows resilient data engineering around messy public-sector sources.
AI review desk	Gemini 2.5 Flash analysis with Groq failover, action-item extraction, evidence quotes, and relevance scoring.	Demonstrates AI that is auditable, not just generative.
Compliance workflow	Review filters, source badges, persisted reviewed state, and document Q&A with page-level citations.	Turns raw circulars into an operator-ready compliance queue.

✨ Demo Story

1. Pull the latest regulatory updates from all five sources.
2. Process each circular into relevance, summary, action items, and evidence.
3. Filter to HIGH, ask questions against the source document, then mark it reviewed.

---

✨ Features

🌐 Multi-source fetch	RBI (RSS) · IFSCA (JSON API) · MCA · SEBI · FATF — parallel, fail-isolated
🎯 AI relevance scoring	HIGH / MEDIUM / LOW / NOT_RELEVANT with Glomopay-specific context
✅ Action items	Concrete compliance steps extracted from each circular
📚 Evidence packs	Verbatim document quotes grounding every AI claim
📄 PDF pipeline	Auto-download → pdf-parse → Tesseract OCR fallback
💬 Document Q&A	Ask any question; get an answer with page-level citations
🔖 Review tracking	Mark as reviewed; persists across sessions (SQLite-backed)
🩺 Source health	Per-source status after each fetch — ok / empty / blocked / error
🛟 Degraded mode	Gemini 2.5 Flash → Groq LLaMA 3.3 failover, surfaced in UI
🧹 Deduplication	INSERT OR IGNORE on URL — zero duplicates, ever

---

🚀 Quick Start

Prerequisites

• Node.js 18+
• GEMINI_API_KEY — get one free
• GROQ_API_KEY — get one free

Install & run

git clone https://github.com/iamadarsha/Assignment.git
cd Assignment
npm install

# Add API keys
cat <<EOF >> .env.local
GEMINI_API_KEY=your_key_here
GROQ_API_KEY=your_key_here
EOF

npm run dev
# → http://localhost:3000

Your first three minutes

1. Fetch Updates → pulls latest circulars from all 5 sources in parallel
2. Process AI → runs Gemini analysis (auto-loops until every circular is scored)
3. Filter by High → review action items for your most urgent obligations
4. Expand any card → Ask This Document to interrogate the full circular with citations

---

🧠 How it works

flowchart LR
    A[5 Regulator Sources<br/>RBI · IFSCA · MCA · SEBI · FATF] --> B[Parallel Fetcher<br/>fail-isolated]
    B --> C[(SQLite<br/>circulars.db)]
    C --> D{PDF?}
    D -->|yes| E[pdf-parse → OCR fallback]
    D -->|no| F[HTML extraction]
    E --> G[Structured Chunks<br/>section + page metadata]
    F --> G
    G --> H[Gemini 2.5 Flash<br/>relevance + actions]
    H -->|fails| I[Groq LLaMA 3.3<br/>fallback]
    H --> J[UI: scored cards + Q&A]
    I --> J

    style A fill:#3b82f6,stroke:#fff,color:#fff
    style H fill:#f59e0b,stroke:#fff,color:#fff
    style I fill:#f59e0b,stroke:#fff,color:#fff
    style J fill:#22c55e,stroke:#fff,color:#fff

Loading

---

🔌 API Reference

Endpoint	Method	Description
/api/circulars	GET	All stored circulars with AI fields
/api/fetch	POST	Scrape all 5 sources, insert new circulars
/api/process	POST	AI-analyse the next batch of unprocessed circulars
/api/chat	POST	{ circularId, question } → answer + evidence
/api/review	POST	{ id, reviewed: bool } → persist review state

---

📁 Project Structure

app/
  page.tsx                  # Main UI — single-page React
  api/
    circulars/route.ts      # GET all circulars
    fetch/route.ts          # POST trigger scrape
    process/route.ts        # POST trigger AI analysis
    chat/route.ts           # POST document Q&A
    review/route.ts         # POST toggle reviewed flag
lib/
  db.ts                     # SQLite schema, migrations, queries
  fetchAll.ts               # Parallel fetcher orchestrator + source health
  fetch{RBI,IFSCA,MCA,SEBI,FATF}.ts   # Per-source scrapers (WAF-aware)
  processCircular.ts        # AI analysis — Gemini → Groq fallback
  chatDocument.ts           # Document Q&A with chunk scoring + evidence
  structureText.ts          # Text → chunks with section/page metadata
  pdfPipeline.ts            # PDF download + parse + OCR
data/
  circulars.db              # SQLite DB (auto-created)
  pdfs/                     # Downloaded PDFs (auto-created)

---

🚢 Deployment

Vercel (one-command)

npx vercel
# Add GEMINI_API_KEY + GROQ_API_KEY in the dashboard

SQLite data won't persist across Vercel deployments (ephemeral filesystem). See ARCHITECTURE.md for the Postgres migration path.

Self-hosted Node.js

npm run build && npm start
# Data persists in ./data/circulars.db — back this up or mount as a volume

Docker

docker build -t regulatory-intel .
docker run -p 3000:3000 \
  -e GEMINI_API_KEY=your_key \
  -e GROQ_API_KEY=your_key \
  -v $(pwd)/data:/app/data \
  regulatory-intel

A fly.toml is also included for Fly.io deploys with persistent volumes.

---

🧰 Tech Stack

Layer	Tooling
Frontend	Next.js 15 · React 19 · TypeScript · Tailwind CSS
Backend	Next.js Route Handlers · Node 18+ runtime
AI	Gemini 2.5 Flash (primary) · Groq LLaMA 3.3 (failover)
Storage	SQLite (better-sqlite3) · file-system PDF cache
Document pipeline	pdf-parse · Tesseract.js (OCR fallback)
Hosting	Vercel-ready · Docker-ready · Fly.io fly.toml included

---

⚠️ Known Limitations

• MCA & SEBI sit behind Akamai/Cloudflare WAFs that block server-side requests. Fetchers degrade gracefully and surface block status in the UI. Production fix: scheduled Playwright job. See HANDOVER.md.
• Single-writer SQLite is sufficient for a prototype; migrate to Postgres for multi-user or concurrent writes.
• Previously-processed circulars keep their old AI output until re-processed. A /api/process?force=true flag is the planned fix.

---

📖 Documentation

File	Contents
ARCHITECTURE.md	System design, data flow, component map
TECH_STACK.md	Every technology choice with rationale
DESIGN_DECISIONS.md	Problem scoping, trade-offs, 3-month roadmap
HANDOVER.md	Developer handover — known issues, env setup, next steps

---

📄 License

MIT — free to use, fork, and adapt.

---

Built by Adarsha Chatterjee · Portfolio →