HakiScribe — Legal work from the room

Privilege-safe legal listening for Kenyan rooms: Nemotron on Nebius turns spoken meetings into source-traced drafts, dates, and research lawyer chooses what leaves the desk.

Comment

Inspiration

After a client meeting in Kenya, the demand letter, the diary date, and the line that must stay private still live in someone’s head. Advocates reconstruct work from memory—or paste a full transcript into a generic chatbot and risk privileged speech. We wanted the next document, not another wall of text: a listening instrument for HakiChain that belongs in the room and keeps liability with the lawyer.

What it does

HakiScribe is a privilege-safe legal work agent:

  1. Listen — phone/laptop mic or Omi wearable (English, Kiswahili, code-switch).
  2. Flag — mark moments without breaking eye contact.
  3. Verify — name speakers; lock privileged lines (visible, struck through, never sent to the model).
  4. Propose — an Action Tray (not a chat): draft, calendar, matter, contact, time entry, note — each with a source quote.
  5. Decide — the professional selects, edits, then generates. Nothing auto-serves as advice. Nebius × NVIDIA: Detect runs on Nemotron Super via Nebius Token Factory; drafts on Ultra; research uses Tavily over Kenyan legal domains, then Super synthesis labelled External authorities (not evidence). When Nebius is connected there is no silent fallback to another provider.

How we built it

  • Frontend: TanStack Start / React (Lovable + Vercel dual deploy).
  • Backend: FastAPI on Render — sessions, privilege filter, Action Tray, Connectors.
  • Models: OpenAI-compatible Nebius Token Factory (api.tokenfactory.nebius.com); live GET /models discovery; Super / Ultra / Nano tier routing.
  • Jobs: Trigger.dev (or in-process) for detect / generate / research.
  • Trust path: redacted TranscriptSegments stripped before every completion; source-quote validation; diagnostics + live demo (/demo/live-nemotron).

Challenges we ran into

  • Making Nebius the real path when older deploy defaults still pointed at OpenRouter—and refusing silent fallback so the demo stays honest for judges.
  • Encoding privilege as a product constraint, not a prompt suggestion.
  • Teaching the UI that Tavily retrieval ≠ verification (clear “not evidence” labelling).
  • Fitting a room → privilege → tray → draft → research story into a sub-3-minute live path.

Accomplishments that we're proud of

  • End-to-end journey: live Nemotron demo → privilege lock → Action Tray → Ultra draft → Tavily research.
  • MIT-licensed public repo judges can clone and smoke-test.
  • Connectors “active tools” surface and backend diagnostics tied to the same settings.

What we learned

Hackathon eligibility means a runtime Token Factory call with an NVIDIA open model—the rest of the stack can stay where it is. For legal users, privilege and source quotes beat model theatre. Honest errors beat a pretty fallback that hides which provider ran.

What's next

Firm tenancy and production auth, measured latency from live smoke, deeper HakiChain matter sync, and more Kenyan-domain research coverage—still with the lawyer choosing what leaves the desk.

Built with (tags — pick up to 25) Nebius · Nebius Token Factory · NVIDIA · NVIDIA Nemotron · Tavily · FastAPI · Python · React · TypeScript · TanStack Start · Vite · Trigger.dev · Render · Lovable · Omi · Whisper · PostgreSQL · WebSocket · OpenAI API · WhatsApp

Try it out links Label URL Live demo https://hakiscribe.lovable.app/ API health https://hakiscribe-backend.onrender.com/health GitHub https://github.com/Isomkevin/haki-scribe Judges: Sign in → Use demo credentials → /new → Run live Nemotron demo (not the saved sample).

Built With

Share this project:

Updates

Submission history