—Applied AI research & products
AI for domains where approximately right is not good enough.
We build intelligent software, AI infrastructure and enterprise systems — while advancing long-term research into the technologies that make those systems trustworthy. Evaluation-first, built to be verified.
- Founded
- 2026
- Pillars
- Products · Research · Engineering · Community
- Default
- Open source
- Based in
- India
01What we do
Most AI projects do not fail because the model is not clever enough. They fail because nobody could say, with evidence, whether the system was working.
That is the gap Rudvanth works in. We build AI systems for organisations that have to answer to a regulator, an auditor, a clinician or a court — where a confident wrong answer is worse than no answer at all.
In practice that means starting from the measurement problem. Before we write the feature we write the evaluation set: the hard cases, the failure modes, the thresholds the system has to clear before anyone is allowed to depend on it. The model is the easy, replaceable part. The harness is the asset.
The company is organised around three pillars that feed each other. Research produces methods. Products turn methods into software we own. Engineering partnerships fund both, and drag the research back to reality every time it drifts.
02How the company is built
Four pillars, one loop.
Each pillar is weak on its own. Research without products is a hobby; products without research are commodities; both without revenue are a countdown; and none of it compounds without engineers who can be persuaded to join.
Products
Systems we build, operate and license ourselves. Each carries a published status label — nothing appears here before it exists.
- Owned IP
- Published status
- Design partners
- No pre-selling
Research
Applied research on evaluation, grounded retrieval, domain adaptation and efficient inference, published as a numbered index anyone can cite.
- Evaluation & assurance
- Grounded retrieval
- Domain adaptation
- Efficient inference
Engineering
We design, build and scale AI-powered software for organisations with real constraints. It funds the research, and it keeps the research honest.
- AI systems & LLM integration
- Voice AI & multilingual
- Enterprise platforms & cloud
- Dedicated engineering teams
Community
Open-source releases, technical writing and hiring in the open. Inspectable artefacts are how technical trust is actually earned.
- Open evaluation harnesses
- Public benchmarks
- Engineering notes
- Coordinated disclosure
03Operating model
The flywheel.
Engineering funds research, research becomes products, products create recurring revenue, and revenue buys the engineers who make the engineering worth more. We would rather draw it than let anyone guess which business they are dealing with.
- 01
Engineering services
Deep delivery work for enterprises
- 02
Sustained cash flow
Unsubsidised, and nobody else sets the clock
- 03
Fund internal R&D
Rudvanth Labs
- 04
Build proprietary products
Intellectual property we own outright
- 05
Recurring revenue
Platforms and licensing
- 06
Hire better engineers
Which raises the ceiling on step 01
↻ back to 01
The order matters. Plenty of companies attempt this loop starting at step four, funded by someone else’s money and a deadline. We start at step one, which is slower and considerably harder to kill.
04Currently building
Four systems in development. No launch dates.
- Evaluation platform
- Cross-industry
- Clinical documentation assurance
- Healthcare
- Obligation intelligence
- Legal & compliance
- Field inspection intelligence
- Manufacturing
05Operating principles
Four commitments we are willing to be held to.
These are not values-page decoration. Each one costs us something specific, which is the only reason it is worth writing down.
Evaluation before capability
A system that cannot be measured cannot be improved, defended, or safely handed to a customer. We build the harness before we build the feature, and we keep the harness in the repository next to the code it grades.
Own the hard part
Wrapping a model API is not a business. We take the work that is genuinely difficult — the retrieval layer, the evaluation set, the domain constraints — and we make that the part we own.
Publish the method
Methods go out in the open even when the product does not. Open evaluation harnesses and honest write-ups cost us nothing we needed to keep, and published methods are how technical trust is actually earned.
Focused teams, long horizons
We would rather do four things properly over three years than twenty things badly in one. Every commitment on this site is one we intend to still be honouring in 2030.
06Where we work
Industries where the cost of being wrong is legible.
We deliberately avoid domains where nobody can tell whether the output was any good. If the mistake is invisible, the evaluation problem is unsolvable — and the work is not interesting.
Healthcare & life sciences
Clinical documentation, coding and prior-authorisation workflows where an error has a named patient attached to it.
Financial services
Underwriting, KYC, surveillance and reporting — domains with an auditor at the end of every decision path.
Manufacturing & mobility
Inspection, maintenance and field intelligence, usually on constrained hardware and intermittent connectivity.
Energy, climate & ESG
Disclosure, measurement and assurance pipelines where the output is a number somebody has to certify.
Public sector & infrastructure
Citizen-facing services in multiple Indian languages, built to be inspected rather than trusted on faith.
Legal & compliance
Contract and obligation analysis where the value is in the citation, not the summary.
07Open source
We release the tools we grade ourselves with.
Inspectable artefacts settle questions that assertions cannot. An evaluation harness someone else can run against their own data says more about how we work than any case study we could write.
$ rudvanth-eval run --suite clinical-coding
suite clinical-coding n=1,284
model under-test temp=0.0
exact-match .................... 0.913
citation-grounded .............. 0.968
refusal-when-unsupported ....... 0.994
p95 latency .................... 1.4s
✓ 4/4 gates passed
→ report written to ./evals/2026-08-07.jsonIllustrative output. Our evaluation tooling and its results are published as they are released.
08Writing
Notes from the work.
What we mean by applied AI
A short statement of what Rudvanth is for, what we are deliberately not going to do, and the trade-offs we are accepting on purpose. Written at the start so it can be held against us later.
Read28 July 2026 · 5 minEvaluation is the product
Every AI system we have inherited had the same defect: nobody could say whether it was working. Here is why we now build the measurement layer before the feature, and what it costs to do it in the other order.
Read14 July 2026 · 5 minRetrieval a compliance officer can sign off on
Enterprise RAG systems mostly fail at retrieval, not generation — and the reasons are structural. Notes on chunking that respects meaning, permissioned search, and why the citation is the product.
ReadStart a conversation
Tell us what has to be right.
If you have an AI system that has to survive an audit, a clinician, or a regulator, we would like to hear about it. We reply to every serious enquiry within two working days.