ComplianceOS: AI Compliance Agent for MSMEs

Inspiration

What it does

Inspiration

India has 64.5 million MSMEs, and a single small business faces 1,450+ compliance obligations a year across tax, labour, environmental, and corporate law. The ecosystem ships 42 regulatory updates every single day. Miss a deadline and the penalties are brutal — MSMEs pay an estimated ₹89,000 crore in penalties annually, almost all from missed deadlines, not wilful evasion. A proper CA retainer costs ₹5 lakh+/year — out of reach for the 5-person shop in Chennai.

We built ComplianceOS so that business owner can have an AI compliance officer for ₹500/month.

What it does

ComplianceOS is an autonomous AI compliance agent. Give it your business profile and it:

Builds your Compliance DNA — matches your state, industry, size, turnover, registrations, and entity type against a regulatory corpus to derive your exact obligation set (a sole prop never sees Pvt-Ltd filings).
Scores every deadline with a Decay Score — a composite urgency score blending time remaining, procedural complexity, penalty severity, and your own filing history, with a dedicated overdue model.
Detects regulatory Ripple — when a new circular drops, a hybrid Vector + Atlas Search retrieval (fused with Reciprocal Rank Fusion) finds exactly which of your obligations are impacted.
Predicts penalties in rupees before they hit, scaled to your size and turnover.
Auto-drafts filing-ready documents and exports real PDFs.
Answers anything through a multi-agent Compliance Officer that streams its reasoning, with a full audit log + replay.

How we built it

Agent brain: Google Gemini 2.5 Flash.
Orchestration: Google ADK — a planner agent routes questions to 6 specialist sub-agents that call deterministic compliance tools, with the reasoning trace streamed live to the UI.
Data layer: MongoDB Atlas, used deeply — Vector Search, Atlas Search, hybrid retrieval (RRF), Time Series collections for decay-score history and exposure forecasting, Change Streams pushed to the browser over SSE, rich aggregation pipelines, and the Atlas MCP Server wired into the agent. Eight distinct Atlas capabilities in production.
Backend: Python 3.11 + FastAPI + Motor (async).
Frontend: React 18 + TypeScript + Vite + Tailwind — a dark "Obsidian + Emerald" command center.
Auth: Firebase. PDFs: ReportLab. Deployed: Vercel (frontend) + Render (backend) + Atlas.

Challenges we ran into

Calibrating the Decay Score. Our first formula divided by complexity and penalty, which pinned every high-stakes filing permanently red. We switched to logarithmic shift terms, giving a smooth, interpretable urgency ranking.
Vector search false positives. Gemini embeddings cluster in a tight similarity band, so pure vector ripple detection over-fired. Fusing it with Atlas Search via RRF — and only calling a hit "direct" when both retrievers agree — fixed the precision.
Keeping it real on free tiers. Graceful degradation everywhere: if an index or the LLM quota is unavailable, the app falls back instead of breaking.

Accomplishments that we're proud of

8 MongoDB Atlas capabilities in genuine production use.
A multi-agent system with a visible, replayable reasoning trail.
A 0-bug, 5-persona end-to-end audit across 40 API routes.
It's live and deployed end to end.

What we learned

How deep MongoDB Atlas goes beyond a document store — vector + lexical hybrid retrieval and Change Streams became the backbone — and how much trust an auditable agent earns versus a black box.

What's next for ComplianceOS

WhatsApp + regional-language alerts, CA collaboration, accounting integrations (Tally/Zoho), and live regulatory-feed ingestion for the whole MSME ecosystem.