Manu
CV
← Projects

Lexora

WhatsApp legal intake platform built in 36 hours at the LLM x Law hackathon. Processes voice notes, documents, and photos, generating structured case files and urgency alerts for legal teams.

Status
Shipped · hackathon
Started
Oct 2026
Stack
TypeScript · Fastify · React · Mistral · Voxtral · Kapso · PostgreSQL · Python
Code
Live demo, 4 Oct 2026: the client talks to the WhatsApp number on the left, the case appears rated Critical on the lawyer dashboard on the right.
Lawyer dashboard (team build, fictional data): open cases sorted by urgency.
Lawyer dashboard (team build, fictional data): open cases sorted by urgency.
Case view: urgency, recommended actions, missing information and the client's documents.
Case view: urgency, recommended actions, missing information and the client's documents.

Context. Built in 36 hours by a team of four at the LLM x Law Hackathon, hosted by Stanford Law School and Mistral AI.

Architecture. A WhatsApp intake assistant for criminal-defense firms:

  • Audio and documents: voice notes transcribed with Voxtral, documents processed through OCR.
  • Governance (Python prototype): pythagoras-sec’s control point and twin’s WhatsApp gateway; every agent action is appended to an HMAC-chained, tamper-evident log.
  • Final build: TypeScript monorepo, where I integrated the Kapso WhatsApp webhooks for the live demo.

Functionality

Clients submit case details to a single WhatsApp number using text, voice recordings, images, or document uploads. Lexora creates a case record, gathers required information, transcribes audio (Voxtral), extracts text from documents (OCR), analyzes image content, queries public legal databases (Légifrance, Judilibre, ECHR), calculates urgency scores, and sends immediate WhatsApp notifications to attorneys for critical cases.

WhatsApp (Kapso) -> POST /webhooks/kapso (HMAC verified)
  -> Case storage -> Intake agent (Mistral) -> Client response
  -> Audio processing (Voxtral) / OCR / Image analysis
  -> Legal lookups -> Urgency score -> Attorney WhatsApp alert
Attorney dashboard <- GET /api/cases

Individual contributions

Team members: Pietro Putelli, Kamil Akerramou, Eliott Floquet, and Emmanuel Noudogbessi.

  • Initial Python prototype: Implemented complete processing pipeline (WhatsApp webhook, Voxtral transcription, OCR, attorney web UI) integrated with the pythagoras-sec control point. Every agent tool call was validated by code rules, personal identifiers were redacted from external queries, and actions were logged to an HMAC-chained audit log.
  • TypeScript production integration: Integrated Kapso WhatsApp webhooks into the final monorepo stack, enabling live end-to-end testing for hackathon judges.
  • Conversational intake agent: Built prompt logic that asks targeted questions based on missing case evidence while strictly refraining from offering legal advice.
  • Audit dashboard: Built attorney UI components displaying full legal search traces, including redacted queries, execution status, and response latency.
  • Deployment runbook: Documented setup procedures, sandbox keys, tunneling configurations, and operational troubleshooting steps for live demonstration.

Results

Completed 36 hour build without official awards. Detailed event review and retrospectives documented in the events section.