Tvaris Labs · Hyderabad
Most AI projects work in a demo and fall over in week three. We build the unglamorous half that decides which way it goes — evaluation harnesses, retries and fallbacks, cost telemetry, and the data pipelines feeding it all.
What we build
Every engagement is some combination of these. If your problem is not on this list, we will tell you so rather than stretch to fit it.
Agents that use tools, hold context across long tasks, and take real actions in real systems — not a chat box in front of a prompt. Built with approval gates where a mistake would be expensive, and containment for when a step fails.
Retrieval over your own corpus, document understanding, and pipelines that turn unstructured input into structured output you can check. Your provider, your keys, your data boundary.
The part most teams skip. Without an evaluation harness you cannot tell whether a prompt change made the system better or quietly worse. We build LLM-as-a-judge suites, regression sets, and token and cost telemetry.
An agent is only as good as what reaches it. This is our deepest bench: large-scale ingestion, transformation, and cost control on Azure and AWS, built at petabyte scale in production.
Selected work
We do not have a wall of client logos. What we have is software we designed, built, shipped, and still operate — billing, authentication, infrastructure, evaluation and all. Every product below is a set of decisions we can walk you through in detail, including the ones we got wrong.
Synthetic data generation for teams who need realistic test data without touching production records. Schema-aware generation, referential integrity across tables, and export into the formats pipelines actually consume.
Python, Streamlit, LLM generation, Azure
datacraves.com →An agentic workspace where language models use tools and hold context across long-running tasks. This is where most of our thinking on orchestration, retries and containment was worked out in production rather than on a whiteboard.
Agent orchestration, tool calling, evaluation harness
robollms.com →The shared substrate under the rest of the portfolio: identity, billing, entitlements and deployment. Built once so each product did not have to reinvent the boring, security-critical parts.
Flask, OAuth, Razorpay, entitlements service
tvaris.com →A focus and study tool built for a real user with a real deadline. Small on purpose, and a useful lesson in how much of a product is the twenty percent people touch every day.
React, TypeScript, Flask API
focustu.com →Machine learning explained in short, structured pieces. Built to test how far generated content can be pushed before it stops being worth reading, and where the editorial line has to sit.
Static generation, LLM content pipeline, evaluation gates
shortml.com →How we work
We take on a small number of engagements at a time, which means we can afford to be honest early about whether yours is one of them.
You describe the problem, the constraints, and what working would actually look like. Free, and there is no form funnel in front of it. If we are the wrong people, you will hear that on this call.
The approach, the milestones, what we will explicitly not do, and the commercials. Scoped against the actual work rather than a seat count, because a retrieval system over a clean corpus and one over twenty years of scanned contracts are not the same project.
Short cycles, each ending in something that runs. You see the work as it happens rather than at a reveal, and you can change direction while changing direction is still cheap.
Documented systems your team can own, or we continue running them. Your choice, and it does not have to be made at the start.
Who we are
There is no account manager layer here. The people you talk to are the people who build the thing.
Co-founder and CEO
Eight years building data and machine learning platforms at scale. At Microsoft, pipelines processing over 200 terabytes of raw logs a day for OneDrive and SharePoint, cutting compute and storage by 70 percent, and an LLM-as-a-judge evaluation harness for OneNote Copilot. Before that, the package lifecycle datasets at Amazon, and four years at JPMorgan Chase.
IIT Madras · Dual Degree, Electrical Engineering
Co-founder and COO
Eight years in product and go-to-market across SaaS, industrial IoT, pharma, and fintech. Currently Head of Commercials at MachineMax, where she rebuilt the go-to-market motion around data rather than instinct. She decides whether what we are building is worth building.
IIM Tiruchirappalli, MBA · NIT Warangal, B.Tech
What we won’t do
Most studio sites only tell you what they are good at. These are the limits, stated before you spend a call finding them.
We have not delivered client engagements yet. Everything in our portfolio is our own software, built and operated by us. We are taking that experience into client work now, and you would be early. That is worth knowing before you read anything else on this page.
We will not put an agent where a script belongs. A good deal of what gets scoped as an AI project is a data problem, a reporting problem, or a process problem. We would rather say that in week one than bill for it in month six.
We do not staff-augment. We are not a body shop and we do not place engineers into your backlog by the month. We take defined problems and deliver working systems.
We do not take on more than we can build. Two people means a small number of engagements at a time. If we are full, we will say so rather than queue you behind a hiring plan.
We do not publish prices. The work varies too much for a number on a page to mean anything. You get a written scope with commercials after one conversation, before any commitment.
A paragraph is enough to start. We reply to everything, and we will tell you honestly whether it is something we should be doing.
Email the studio