MikeRoss

Inspiration

Access to justice is expensive and slow. Friends facing urgent legal issues got wrong answers from generic chatbots. We set out to build a system that reads the law, cites it verbatim, and routes users to the right tools or professionals.

What it does

• Answers legal questions with cited passages from official statutes.
• Retrieves official forms (e.g., Ontario LTB N5) and lawyer referral options.
• Keeps conversations stateful and structured, never “winging it.”
• Generates clean PDFs from the final guidance.

How we built it

Tech Stack

• Frontend: Next.js 14 + Tailwind (RSC, dark premium UI).
• Backend: FastAPI with CORS for Vercel/Edge.
• Orchestration: LangGraph state machine (Router → Research → Generator).
• Model: Google Gemini 2.0 Flash with structured output (Pydantic schemas).
• Embeddings: text-embedding-004 (768‑dim).
• Memory: 16,170 official legal document chunks in MongoDB Atlas Vector Search (HNSW).
• Querying: Jurisdiction pre-filter → vector search → synthesis with citations.
• Endpoints: POST /chat (thread_id), POST /generate-pdf, GET /health.
• Infra: Vercel (frontend), Railway (backend), MongoDB Atlas (data).

Architecture

• Router classifies intent, topic, jurisdiction.
• Research runs Atlas similarity_search with pre_filter to eliminate cross‑jurisdiction noise.
• Generator returns structured JSON: explanation, citations, options (buttons for forms/referrals).
• Tools: Official Form Finder and Lawyer Referral integrated into the graph

Challenges we ran into

• Production DB/TLS connectivity and IP allowlisting.
• Consistent npm/python tooling across machines (lockfiles, pyenv).
• Designing prompts that force citations and avoid hallucinations.
• Jurisdiction bleed in vector search — solved via strict pre_filter by jurisdiction, better chunking/metadata, and tuned HNSW params. -Reliable structured outputs with citations — enforced Pydantic schemas, validators, and “cite-only-from-context” prompts.

Accomplishments that we're proud of

• End‑to‑end deployed demo with sub‑second retrieval (~0.4s).
• Deterministic state graph routing and structured outputs.
• Clean UX with citation rendering and PDF export. ## What we learned
• LangGraph enables reliable multi‑step reasoning vs. linear chains.
• MongoDB Atlas HNSW with pre‑filters beats naive RAG on jurisdictional precision.
• Strict output schemas are key to trust in legal contexts ## What's next for MikeRoss
• “Saul Goodman” educational mode.
• Voice calling workflow (“Call my lawyer”).
• DocuSign-style auto‑fill for retrieved PDFs.

Built With

• fastapi
• google-gemini-2.0-flash
• langgraph
• mongodb-atlas-vector-search
• next.js
• pydantic
• python-3.11
• railway
• tailwind-css
• text-embedding-004
• vercel