It is easier to describe a legal AI product by what it refuses to do. So: three commitments, and one refusal.
Cited, or it does not ship
Every substantive answer resolves to a passage in a real document. Not a plausible summary of the area — a paragraph, in a judgment, that you can open.
Your dates, not generic ones
A limitation warning that does not know your accrual date is a horoscope. The practice side exists so the research side has something real to reason about.
What that means in practice
Matters, hearings and tasks are first-class, not a CRM bolted on.
Cause-list sync is continuous, because the data changes hourly.
Every derived date shows its working.
All 22 languages, properly
Not translation as a wrapper. The corpus is multilingual, which is a different and much harder claim.
The refusal
We will not give the product a view. It will not tell you whether to settle, whether your case is strong, or what a bench is likely to do.
A tool that has an opinion is a tool you have to argue with. A tool that has evidence is one you can use.
This was once on the roadmap. It is not any more, and removing it was the best scoping decision we made this year.
Isn't outcome prediction the obvious feature?
It is the obvious demo. It is a poor feature: the training signal is thin, the failure mode is confident and invisible, and the professional consequence of being wrong lands on the advocate rather than on us.
Benchmarks on how litigation actually behaves — durations, distributions, most-litigated provisions — are useful and honest. A number attached to your matter is neither.
A short demonstration
The embed below exists to prove the renderer handles one; it is a placeholder, not a product video.
What is next
Cited research across the Supreme Court corpus
Matters, hearings and limitation
Drafting in court format, from your own matter
The India Legal Trends report