AGI Labs · Applied AI studio · Kumbakonam, India · Est. 2026
AGI Labs builds AI products for domains where a wrong answer has a cost — law first.
Our flagship is Tark, a rehearsal courtroom for Indian advocates: argue your matter against an AI opposing counsel, before an AI bench, and get a scored review before you stand up in a real court. Its arguments are built from retrieved judgments, not invented.
Flagship · Tark
An advocate's first hard questions from a bench usually arrive in a live hearing. Moot courts help, but they need a room full of people and they don't scale. Tark puts the room in software: one advocate, an opposing counsel, a bench — each role played by Claude, each grounded in Supreme Court of India precedent.
Set out the facts, the issues and the side you're arguing. Tark builds the brief for the other side from the same record.
Opposing counsel rebuts with cited precedent. The bench interrupts, asks for authority, and presses on the weak point — the way a real bench does.
A scored assessment of the hearing: which arguments held, which authorities you missed, and the judgments you should have cited — each one linked to its source.
Who it's for: litigating advocates preparing a hearing, junior advocates building courtroom judgment, and law students preparing for moots.
The record
A courtroom simulation is only as honest as its authorities. Before writing a single persona, we built the legal data layer Tark argues from: the Supreme Court corpus, a citation graph across it, and an evaluation harness to measure retrieval before anything is claimed.
Every citation in every judgment is parsed, resolved to its target and name-checked, so precedent is a graph we can traverse — which authorities a judgment rests on, and which later judgments rely on it.
A 3,000-query benchmark with a held-out test split. Using the citation graph as a prior on top of passage-level BM25 lifted MRR@10 from 0.138 to 0.183 on the test set — about a third better at putting the right judgment near the top.
Also from the studio
Tark is the product we're taking to market. These are the systems where we proved the approach — an assistant that retrieves from a real corpus and can show you the line it used.
A WhatsApp system for an Ayurvedic practice, used with real patients. It greeted each one in their own language, collected what the doctor needed, and handed over to a human where a machine shouldn't be answering.
All 701 verses with audio, word-by-word meanings and classical commentaries. Ask in plain language and get the verse it came from, not a paraphrase of one.
1,556 pages of classical texts indexed, with the source layer kept strictly separate from corrections and derived rules — so a claim can always be traced back to its page.
CV analysis and structured self-appraisal with a graph as memory: what the agent learns persists as nodes and edges it can query later.
How we build
Founder
Founder, AGI Labs
AI engineer who has spent his career shipping production retrieval and agent systems: a 65,000-node parts knowledge graph for a Fortune 500 manufacturer at Karini AI, GraphRAG for product data at Mercedes-Benz R&D India, and now an agentic trading backend at Neo Wealth. M.Tech in Data Science & Analytics, IIIT Prayagraj.
Contact
Advocates and chambers who want early access to Tark, law schools interested in moot preparation, and partners with a corpus that deserves a proper assistant.
adithyagiridharan@agilabs.co.in