Skip
// AI·2025

OFICINA AI

RAG assistant for a firm sitting on 1.2M documents.

ClientInternal · LegalTech client
Year2025
CategoryAI
Duration20 semanas / 20 weeks
RoleML engineering, full-stack, evaluation

// Context

A 240-attorney law firm wanted a copilot that answers with verifiable citations from their internal case-law corpus.

// Challenge

Guarantee citation accuracy with no hallucinations, per-matter authorization and < 4s latency on answers spanning 8+ relevant docs.

// Approach

Ingest pipeline with OCR, embeddings on pgvector, hybrid BM25 + dense retriever, generation with Claude, and automated eval on LangSmith. UI with inline citation verification.

// Stack

  • Claude API
  • LangChain
  • pgvector
  • Postgres
  • Tesseract
  • Next.js

// Results

0sp95 latency
0%citation precision
0hsaved per attorney/week