Insafdaar

The interesting part is what happens between the CRUD screens. Every case walks a 13-stage lifecycle modeled as a state machine; no case ever lands in the wrong stage.

Across three portals (client, advocate, admin), users get JWT auth with role-based access, OTP-based e-contract signing, and PDF invoice/voucher generation with payment verification. OCR extracts documents, advocate matching ranks lawyers by fit, and Google Calendar and Meet plug in for scheduling. The React SPA with Tailwind CSS is bilingual (English/Urdu via react-i18next), served behind Nginx, deployed on Oracle Cloud with Docker Compose and CI/CD via GitHub Actions.

Three standalone AI microservices, all FastAPI on Cloud Run, plug in via REST APIs and webhooks:

• AI Document Drafting: generates 9 types of legal documents (plaints, affidavits, appeals, contracts) with Gemini 2.5 Flash on Vertex AI; it pulls case context from the database, runs it through a multi-stage generation pipeline, and exports formatted DOCX/PDF. Supports section-level regeneration with custom advocate instructions.

• Voice Intake Agent: conducts bilingual (Urdu/English) intake interviews over WebSocket audio, with GCP Speech-to-Text for real-time transcription and Gemini 2.5 Flash for conversation flow and entity extraction; structured JSON out with parties, locations, dates, amounts. Audio stays in GCS for 7 days.

• Legal RAG Assistant: self-hosted Weaviate (v4) on GCE; ingests Pakistani case law and CPC section PDFs from Google Drive, chunks and embeds them (Gemini text-embedding-005), and answers legal questions with cited sources via hybrid search and Gemini 2.5 Flash.