FIR Saathi

• 

  The product promise, human-review boundary, and current evidence signals.

• 

  The real intake screen: ten language choices, source-first wording, and Experimental labels for the seven personally unverified options.

Inspiration

I was inspired by a simple but serious problem: a person's entire account of a crime can be reduced to a few formal lines during FIR intake. Language barriers, illiteracy, disability, and time pressure can make this worse. I wanted to build something that makes the complainant's own voice the starting point of the record.

What it does

FIR Saathi lets a complainant speak their story in English, Hindi, or Gujarati. It preserves the verbatim statement, transcribes it, identifies missing legally relevant details, asks targeted follow-up questions, and generates a structured FIR draft. The draft is read back for confirmation, then a constable verifies every field before signing. Corrections are logged for accountability.

How I built it

FIR Saathi is a browser-based application built with Cloudflare Pages and Workers, Supabase, Groq's Whisper and gpt-oss-20b, and strict JSON-schema constrained generation. Audio is encrypted in-browser with AES-GCM-256 and fingerprinted with SHA-256. A hand-verified BNS 2023 reference list grounds offence suggestions. I designed the system for future Raspberry Pi 5 + Hailo-8L edge deployment.

Challenges I ran into

The hardest challenge was preventing AI from "improving" someone's story and accidentally changing who did what. I also had to handle code-mixed speech, legal terminology, unreliable offence classification, privacy, and making the entire workflow usable for someone who cannot read or type.

Accomplishments that I'm proud of

I'm proud that FIR Saathi became a working, publicly accessible test deployment rather than just a concept. I built the complete pipeline from voice capture to verified FIR generation, including encryption, audit trails, accessibility features, statutory grounding, and human verification — as a 14-year-old student.

What I learned

I learned that building AI for high-stakes situations isn't just about making the model smarter. Constraints, transparency, human oversight, data handling, accessibility, and knowing when the AI should say "REVIEW" are just as important.

What's next for FIR Saathi

Next, I want to strengthen key custody, expand language and dialect support, improve on-device processing for low-connectivity stations, and test the system with real users and domain experts. My long-term goal is to build a system that can assist police stations without ever replacing human responsibility.

Built With

• 5
• accessibility
• ai
• engineering
• generative
• groq
• hailo-8l
• human-in-the-loop
• language
• multimodal
• natural
• pi
• postgresql
• privacy
• processing
• raspberry
• recognition
• supabase
• text-to-speech
• web
• whisper