Every answer carries the judgment it came from.
The legal companion for Indian advocates: research a matter in plain language, draft the filing, and read the judge's exact paragraphs behind every sentence — before you rely on it.
No sign-up, nothing to install. Research aid only — not legal advice.
If a statement can't be traced to a judgment, it is not shown.
Tap any citation and read the exact paragraphs, in-line.
A case is never labelled good law until that is proven.
Question text is not persisted server-side.
Four capabilities. Two are live today.
Legal research
Describe the matter the way you would to a colleague. Get a reasoned answer where every sentence is pinned to a judgment paragraph, with court, date and bench on each authority.
Drafting a filing
Bail applications, s.138 demand notices, writ petitions, plaints and appeal memos — built only from the facts you give, with missing particulars flagged rather than invented.
Argument preparation
Issue-wise argument outlines with the authority for each limb, the counter you should expect, and the weakest link in your own chain.
Courtroom simulation
Rehearse before you appear: bench-style questioning on your own draft, opposing counsel pushing back with citations, and a transcript of where you were pinned.
From judgment to answer, in six steps you can audit.
Court judgments from the public Indian court corpora, cleaned to machine-readable text.
Reasoned judgments kept; adjournments and listing orders held out of retrieval.
Each judgment split by role — facts, issues, arguments, reasoning, ruling.
Citations between cases become a graph: followed, distinguished, overruled.
Your matter is classified to an area of law, then search is scoped to it.
Every sentence must carry a citation, or it does not reach you.
No citation, no sentence.
An assistant that sounds confident is worthless to an advocate; one that shows its authority is a colleague. Answers are assembled only from the judgment collection — never improvised from general knowledge — and every claim is tappable back to the paragraph it rests on. Where the record is silent, the assistant says so and asks you for the missing fact.
Answers and drafts alike close with the authorities they were built from.
Precedent status stays treatment unverified until validation and human legal review clear it.
Plain-language matter descriptions in, filing-shaped output out.
Put it on a matter you are researching this week.
Open the assistant, ask in your own words, and tap through the citations. If a case looks wrong, one button tells us.
Answer the question you were actually asked — with the judgments to back it.
Describe the matter: parties, facts, stage, jurisdiction. The assistant classifies the area of law, searches only within it, and returns a reasoned answer in which every sentence is pinned to the paragraph a judge wrote.
What you get back
Written in continuous prose, addressed to your facts, with the citation carried inside each sentence.
Every citation chip opens the exact passage — role, paragraph numbers and text — without leaving the answer.
Court, decision date, bench strength, posture and area of law on every authority returned.
If a critical fact is absent, you are asked for it instead of handed a half-answer.
“Both the Courts below, on the strength of the admissions made in the counter filed in the rent control proceedings, had come to a categorical conclusion that the petitioner's father had entered into possession of the premises as a tenant under the respondents.”
Real passage from the pilot collection. You read the court's words, not a paraphrase.
Coverage
The pilot collection spans the Supreme Court of India and High Court matters across criminal law, regular and anticipatory bail, cheque dishonour, rent control and property, service, writ jurisdiction and tax. Procedural orders — adjournments, listings, routine directions — are deliberately excluded, so what you search is reasoned law. Courts and periods are being extended continuously; the assistant answers only from what it actually holds.
A first draft you can edit and sign — never one that invents your facts.
Pick the filing, state the matter facts, name the court. You get a structured, editable draft with the authorities it relied on, and an explicit list of every particular the assistant refused to guess.
Filing types available today
Sections invoked, custody status and stage drive the grounds.
Built around the record you supply, with authority for each ground.
Statutory particulars and limitation dates surfaced for your confirmation.
Grounds framed against the impugned action you describe.
Cause of action, relief and valuation laid out for completion.
Grounds keyed to the findings you are challenging.
Discipline built into the draft
Names, dates, amounts and sections come only from your facts. Anything absent is listed as “not specified” for you to fill before filing.
Each draft closes with the judgments relied on, tappable to the paragraph, so you can verify the ground before you sign.
The draft opens in an editable panel; copy it out or download it as text and take it into your own template.
Argument preparation
Walk into the hearing knowing which limb of your case is weakest — and what the other side will reach for. Argument preparation turns your matter into an issue-wise outline, each limb carrying its authority and its expected counter.
What it will do
Your facts broken into the issues the court will decide, in the order it will decide them, with the authority for each.
For every proposition, the judgments the other side would cite against it — including cases that distinguish yours.
Which limb rests on the thinnest authority, and what evidence or pleading would shore it up.
Roadmap
Indicative sequencing, shaped by pilot feedback.
The retrieval and citation-enforcement engine that argument preparation is built on is already running in the pilot.
Outline generation with an authority per limb, released to pilot advocates for correction.
Treatment relationships between judgments — followed, distinguished, overruled — used to surface the opposing authority.
Outlines saved per matter, shared across a team, versioned alongside the draft they support.
Pilot advocates get first access — and shape what ships.
Courtroom simulation
Rehearse the hearing before it happens. The bench questions your own submission, opposing counsel pushes back with real authority, and you leave with a transcript of every point where you were pinned.
What it will do
Questions put to your submission in the register a court actually uses — on jurisdiction, maintainability, limitation and the record.
Objections grounded in judgments from the collection, so the pushback is real law and not invented resistance.
A record of the exchange, marking where an answer had no authority behind it and what to read before you appear.
Roadmap
Indicative sequencing; the simulation follows argument preparation.
Every simulated objection will draw on the same cited-judgment engine serving research today.
Upload or paste a submission and take questions on it, with a small group of pilot advocates rating realism.
Adversarial objections drawn from the judgment collection, each one tappable to the paragraph it rests on.
Every session recorded, with the unsupported answers flagged and reading listed against each.
Repeatable sessions for juniors and students, on real matters and reported facts alike.
Tell us the kind of hearing you would want to rehearse.
The grounding pipeline, end to end.
An advocate should not have to trust a black box. This is what happens between a judgment being published and a sentence appearing in your answer.
Ingestion and cleaning
Judgments are taken from the public Indian court corpora and converted to structured text: headers, footers and watermarks stripped, hyphenation repaired, paragraph numbers preserved, and separate opinions — majority, concurring, dissenting — detected as distinct texts. Paragraph numbers survive the whole pipeline, which is what makes a citation checkable.
Triage: reasoning versus routine
Most court documents decide nothing — adjournments, listing directions, routine disposals. Each document is classified as substantive or procedural, and only reasoned judgments enter primary retrieval. That is why results are authorities rather than paperwork.
Rhetorical-role segmentation
Every judgment is segmented the way a lawyer reads one: facts, issues, arguments, reasoning, ruling. Retrieval can then prefer the part of the judgment that actually answers your question — the reasoning, not the recital of facts — and each retrieved passage arrives labelled with its role and paragraph numbers.
Domain classification
Each judgment carries one or more area-of-law labels — bail, cheque dishonour, rent control, motor accident, partition, service, tax, quashing and more. Your matter is classified the same way, so search happens inside the right body of law instead of across everything at once.
Citation graph and precedent status
Citations between judgments are extracted and classified by how the later court treated the earlier one — followed, distinguished, overruled — with a confidence score. Status is displayed conservatively: until a claim clears strict validation and human legal review, a case is shown as treatment unverified rather than asserted to be good law.
Hybrid retrieval and citation-enforced synthesis
Semantic search over passages is combined with graph traversal over the citation network, scoped to your area of law. The answer is then written under one hard constraint: every sentence must carry the identifier of a retrieved passage. Sentences that fail the check are not shown to you — which is also why the assistant will ask for a missing fact rather than fill the gap itself.
Privacy
Your question text is not persisted server-side beyond transient processing. Connection settings you enter stay in your own browser. Feedback you submit records the answer identifier, not your matter.
Use it, then tell us where it fell short.
The assistant is open — no gate, no sign-up. If you want early access to argument preparation or the courtroom simulation, or you are evaluating this for a chamber or firm, write to us.
Write to us
Email is the fastest route. Tell us the area of law you practise in and what you were trying to research.
[email protected]GradientDescent AI · Legal Companion. A research aid for advocates — not legal advice.
Start without us
Ask a real question, tap through the citations, press the thumbs.
Every “report wrong status” is treated as a priority defect and investigated.
Coverage is extended in the order advocates ask for it.
The assistant
Problem statement → cited judgments with precedent status.
Held in this browser only. Your question text is never persisted by the service.
{{ item.qLabel }}
{{ ch.text }}
Grounded answers only: every sentence carries a citation to the judge's exact paragraphs. Status badges read verified good law, treatment unverified or negative treatment — today the assistant labels cases conservatively as treatment unverified. Nothing you type is stored beyond transient processing.