AdalatAI — AI-Assisted Court Workflow Where the AI Never Convicts and Never Sentences

---

title: "AdalatAI \u2014 AI-Assisted Court Workflow Where the AI Never Convicts and Never Sentences" published: true description: "Bangla legal-tech that quotes statutes verbatim, assembles judgment drafts from verified parts, and leaves the decision to a human judge \u2014 with a full audit trail." tags: ai, bangladesh, showdev, opensource

The problem

4,742,731 pending cases (Supreme Court data, Dec 2024) — roughly half a century to clear at current rates. Records have been digitized since eCourt launched in 2009, but understanding hasn't been: litigants can't easily read their own case status, and no AI assists the workflow safely.

Legal AI has a trust cliff: a wrong-but-confident answer is worse than no answer. So AdalatAI is built around a different principle:

AI never convicts and never sentences. It suggests and cites; the human judge decides and signs — and every suggestion is auditable forever.

Try it: rudra496.github.io/adalatai

What it does

• 🗂 Case intake — narrate facts in Bangla or English (voice supported) → AI suggests case types and applicable Penal Code sections with verbatim statute quotes and punishment ranges parsed from the official text
• ⚖ Judge console — review every AI suggestion → edit / approve / override → sign; complete audit trail
• 📜 Judgment draft assembler — drafts are assembled only from verified parts (facts + verbatim statute + statutory range); the judge edits and signs
• 📚 Law browser — search all 547 sections (official bdlaws numbering), verbatim, with source links
• 📊 Pendency analytics — verified backlog numbers with sources

The zero-invention guarantees

This is the part I care about most, so it's engineered, not promised:

• Statute text is verbatim from bdlaws.minlaw.gov.bd (the official Ministry of Law site), fetched and integrity-tested — there are automated statute-verbatim gates in the test suite that fail if any AI output paraphrases the law
• Classifier: held-out micro-F1 0.997 on a disclosed synthetic corpus, with per-class thresholds calibrated on hand-written natural FIR sentences
• Retrieval: BM25, recall@5 = 1.0 through the real pipeline
• Evidence: 100 Crossref-verified papers (docs/EVIDENCE.md) plus a 100-entry dataset registry with live URL checks
• 18 automated tests, including the verbatim gates; the improve-loop (train → evaluate → critic → remediate) is traced in case_model.json

Why deterministic-first

The rule layer (hand-crafted Bangla/Banglish/English crime-rule patterns) is primary, and the trained classifier only fills gaps with calibrated thresholds. In legal work, recall on the rule layer is auditable line-by-line — a black box isn't.

Links

• Live: rudra496.github.io/adalatai
• Source: github.com/rudra496/adalatai

If you're working on legal tech for South Asia — or on making legal AI auditable — I'd love to compare notes.