#58
Monthly Rank
#160
Lifetime Rank
Share
10 September 2026
Recourse

Recourse

The legal AI that isn't allowed to make up the law. Access-to-justice help for the citizen — every section and judgment retrieved and quoted, never generated. Live at recourse.co.in, and on WhatsApp

Tanbir Yusuf AhmedGitHubWebsiteVideo
Open Source

Gallery

About the Project

In India, the people most exposed to the law have the least access to it — someone arrested at night, or the family standing outside the police station. Most cannot reach a lawyer at the moment it matters. And since July 2024 the criminal codes changed wholesale — the IPC, CrPC and Evidence Act replaced — so even the sections people half-remember are now wrong.

A general-purpose legal chatbot is dangerous here: it gives a confident answer and cites a judgment that does not exist. In a bail matter, that costs someone their liberty.

It is an access-to-justice tool for exactly that moment — built for the person the case is happening to and their family.

The design is inverted. You describe what happened in plain words. An AI reads your situation and writes the explanation — but it does not decide the law. Every section and judgment is retrieved from a curated, human-checked corpus and shown to you word-for-word; the finding on each arrest safeguard, and the draft petition built from your own facts, are produced by fixed, deterministic rules.

It runs today at recourse.co.in, covering arrest and FIR, custodial violence, default bail, Section 138 cheque cases, frozen bank accounts and cyber-impersonation. For a family, it is legal first aid. For a legal-aid lawyer or a junior on bail work, it is a** Monday-morning tool** — paste the client's account, get the safeguard breaches mapped to the new BNSS sections with the governing judgments, and a first draft of the bail or quashing petition in minutes.

The same engine also runs on WhatsApp — no app, no website, just a message, because the people this is built for often reach for their phone before they reach for a browser. Ask a follow-up and it remembers what you already told it, the way a person would. Proven on real conversations; a public number is the very next step.

Built solo with AI-assisted technology — used, pointedly, to engineer the one thing a legal AI must get right: knowing where the model cannot be trusted.

Practice Areas

Key Features

The AI is kept out of the legal reasoning. The model reads your situation and writes the explanation — but the sections, the judgments and the safeguard findings are retrieved and quoted, or computed by fixed rules. Never generated by the model. No invented case law, by construction.

Every answer traced to source, shown verbatim. The exact BNS/BNSS sections and the governing Supreme Court / High Court judgments, each as a word-for-word passage — never paraphrased into existence. "Read the source" shows only what the answer actually used.

Plain-English intake, across the moments that matter. Describe what happened the way you'd tell a friend — arrest and FIR, custodial violence, default bail, cheque bounce, a frozen bank account, cyber-impersonation — and get what the law requires and the concrete first steps.

Arrest-safeguard check, fully deterministic. Nine plain yes/no questions produce a finding on each safeguard (written grounds, 24-hour production, night arrest of a woman, default bail, medical exam…), each with its section and its judgment, and a severity band — "Critical — rights likely violated" down to "No safeguard clearly broken".

Document check. Upload the arrest memo, FIR copy or remand order; Recourse tests each safeguard against the actual paper, with fixed rules.

Draft petition → editable → PDF. A High Court petition assembled from your own facts — cause title, grounds, prayer, affidavit, index — editable on the page and one click to a formatted PDF. For illegal arrest, Section 138 cheque cases and frozen accounts.

Old law → new law, automatically. 1,000+ IPC→BNS and CrPC→BNSS section mappings — the concordance every Indian litigator has needed since the codes changed in 2024.

The same engine, on WhatsApp. Identical safety architecture — no app, no login, just a message. Follow-up questions are understood in context, without repeating the facts. A slow-network failure that once produced the same answer three times was found on real traffic and fixed with an instant acknowledgement and message-level deduplication. Proven end-to-end; a public number is next.

Live, and honest about its limits. Working at recourse.co.in, with test suites behind the deterministic layers; where the corpus is thin, it says so instead of guessing, but it is growing every single day.

Help Needed

Recourse is an ongoing project. Three things would move it forward fastest:

Advocates and legal-aid practitioners to pressure-test the answers and grow the verified judgment corpus — especially DLSA panel lawyers and anyone doing bail and remand work.

Feedback on the petition templates from litigators who file criminal writs and quashing petitions, so the drafts need less editing before they're usable.

Help extending coverage to the next everyday matters — maintenance, domestic violence, tenancy and eviction, motor-accident claims, and more of the IT Act.

WhatsApp pilot testers. The engine now runs as a WhatsApp assistant too, built to reach a family in the moment they need it rather than after they think to search for a website. It's in sandbox testing now — a few real conversations would help most before opening a public number.

If you work in criminal legal aid or run a clinic and would try it on real matters, I'd like to hear from you.

About the Creator

TA
Tanbir Yusuf Ahmed
Former Executive Magistrate . Building legal technology

Building tools that make procedural safeguards usable by the people they protect.

LinkedIn