Skip to content

Applied AI infrastructure

Speech and reasoning, self-hosted.

PunkRaven builds self-hosted AI infrastructure: a speech layer that transcribes and translates all 22 scheduled Indian languages, and a reasoning layer that grounds every claim in a real, retrieved source. That reasoning layer attaches a citation to what it retrieves, and abstains when the sources do not support an answer. Alongside the two layers we are building one application of our own, LawSafe, a separate product rather than something stacked on top of them. Everything runs inside your own network, it is pre-launch, and it says plainly when it cannot verify what it is about to tell you.

PunkRaven builds software. It does not provide legal services and is not a law firm.

See what we are building

Most software built for India was built somewhere else and pointed at India.

The models handle a few of our languages well and treat the rest as an edge case. Systems trained on other countries' material do not know our domains, and cannot tell you which parts of what they know have gone stale. The infrastructure sits offshore on per-call pricing no Indian buyer can renegotiate. And when these systems do not know something, they rarely say so - they produce something fluent and wrong, which in a compliance record, a medical note or a legal matter is worse than producing nothing at all.

  • Our languages are the long tail somewhere else

    The largest Indian languages get decent coverage. The other twelve scheduled languages get an apology. If your users speak Santali, Kashmiri, Manipuri or Bodo, you are building for people the market has decided to round down.

  • Fluent and wrong is the default failure

    A misheard word gets translated confidently into something that was never said. A fabricated reference arrives correctly formatted. Neither surfaces as an error, and most systems will not tell you which parts they were unsure about.

  • A metered offshore API is a cost that never stops

    Per-call pricing to an offshore vendor scales against every product that grows, the weights are never yours, and your users' most sensitive material leaves the country to be processed.

Two layers of infrastructure. One application alongside them.

PunkRaven builds the parts of the stack that are hard to buy and harder to fake: the layer that understands Indian speech, and the layer that reasons over authoritative material without inventing any. Both are products in their own right, with buyers who have nothing in common. Alongside them we are building one application of our own. LawMan reasons; LawSafe is what a person opens.

The infrastructure

  • TNT - the language layer

    Status: Planning

    Audio in, transcript and translation out, across all 22 scheduled Indian languages, through one self-hosted API. MIT-licensed weights, so nothing leaves your infrastructure.

  • LawMan - the reasoning layer

    Status: Specified

    Retrieves the governing material first, answers from what it found, verifies every reference against the source text, and abstains when the sources do not support an answer. Built for Indian law as its first body of authority.

Alongside it

  • LawSafe - the first application

    Status: In design

    A chat-first way for any Indian to describe a legal problem in their own language and understand where they stand, then reach a verified advocate when a human is genuinely needed.

Law came first because it is the hardest available test of a grounded system: the sources are authoritative, the language is exact, and a fluent invention is not a rough draft but a liability. Getting it right there is what makes the rest credible. The same two layers serve any domain where being confidently wrong is expensive.

  • TNT

    Status: Planning

    Audio in. Transcript and translation out.

    A self-hosted speech pipeline for all 22 scheduled Indian languages. One API call returns a punctuated transcript, a translation, and an honest confidence score for both. MIT-licensed models, so nothing leaves your infrastructure and there is no per-call vendor fee.

    Its buyers are contact centres, consumer and social apps, government and public services, media and education - any team whose users do not speak English.

    22 scheduled languages - MIT-licensed weights, self-hosted

    Explore TNT

  • LawMan

    Status: Specified

    AI built for Indian law - that shows its source, or admits it has none.

    LawMan researches statutes, case law, and current rules and procedure, and grounds every legal claim in real retrieved material. References are checked before you see them. When the sources do not support an answer, LawMan says so instead of inventing one.

    Works in Indian languages, by voice or text. Reads real legal documents. Runs on infrastructure you control.

    Explore LawMan

  • LawSafe - the application layer

    Status: In design

    The first application we are building.

    Most Indians who are wronged never reach a courtroom, and often should not have to. They do not know their rights, do not know free legal aid exists, and cannot tell whether their problem is even a legal one. LawSafe lets a person describe what happened in plain language - typed or spoken, in their own language - and get a grounded, cited explanation of where they stand and what their options are. When the matter genuinely needs a human, it connects them to a Bar Council-verified advocate who specialises in exactly that issue.

    Understanding first, transaction second - free at the first mile, in the language the person actually speaks.

    Explore LawSafe

The same six rules, in every system we ship.

These are not values on a wall. Each one is a constraint that shows up in the architecture, and each one costs us something we have decided to pay.

  1. Grounded, or silent.

    Our systems answer from retrieved current material and attribute the claim to its source. When the sources do not support an answer, the system says so instead of inventing one. A wrong answer is worse than no answer, and we would rather ship a product that sometimes declines than one that is never uncertain.

  2. Confidence is part of the contract, not a footnote.

    Uncertainty is returned to you, not smoothed over. TNT scores recognition and translation separately for every segment and states which quality tier the language sits in. LawMan treats an uncited claim as a defect rather than a stylistic choice. We would rather tell you a language is harder than quietly pretend all 22 are equivalent.

  3. Runs where your data is allowed to be.

    MIT-licensed weights and single-unit deployment mean our systems can run entirely inside your network. No audio leaves your infrastructure. No confidential document reaches a third party. No vendor can change per-call pricing under you.

  4. Built for the language, not translated into it.

    All 22 scheduled languages, not the ten that are comfortable. Domain structure and convention learned properly rather than a foreign model asked politely about them. Scanned documents at the print quality Indian institutions actually produce.

  5. Skill in the model, facts in the sources.

    Our systems are built to reason, to use domain language correctly and to cite properly. The facts always come from live retrieved material. That separation is why the knowledge does not go stale - and why the domain is a parameter rather than something welded into the model.

  6. We say what we have not built yet.

    Every performance and cost figure we publish is labelled as an estimate or a measurement. There are no customer logos on this site, because there are no customers yet. A company whose products are built to admit uncertainty does not get to overclaim on its own homepage.

Early, and saying so.

PunkRaven is pre-launch. TNT has a complete technical specification and is at planning stage. LawMan is fully specified and not yet built. LawSafe is in design. Nothing on this site is a description of running production software, and where we have given a number it is an engineering estimate for planning rather than a benchmarked result.

Where each project stands
ProjectStageWhat exists today
TNTStatus: PlanningFull technical specification and costed deployment plan
LawManStatus: SpecifiedComplete system design, pending build
LawSafeStatus: In designProduct vision and scope