Mostly in healthcare, legal and government work — where records are private and answers get checked. I design how the data stays protected, how the model reasons, and how every answer can be traced back to its source.
Shipped for
Ajith Thaduri
AI Engineer · Consultant
PHI
18+ categories redacted before any reasoning
Deterministic
Same case in, same answer out
Verified
Every finding reproduced before it's reported
Selected work
Each one is described by the problem it had to solve. Client names stay private; the thinking behind each is on its own page.
Smaller projects and engagements, across the same kind of work.
Data-sensitive enterprise
Adapted an open-weights model to the client's domain, compressed it to fit their existing GPUs and served it inside their own network, so no request ever left. The project existed because a hosted API wasn't an option.
Customer support
Speech in, reasoning and tool calls in the middle, ElevenLabs speech out. Most of the engineering is about timing — a reply that arrives a beat late stops feeling like a conversation — so the pipeline is built around latency.
Product engineering teams
A domain eval set with automated scoring, wired into the delivery pipeline, so a prompt change, model swap or new quantization can't ship if it makes the task worse. It turned “it feels better” into a number the team could discuss.
Blueprints
How I'd design common AI systems — retrieval over long or sensitive documents, agents, guardrails — written up in enough detail to use. Free to use and adapt; a link back is appreciated.
How I work
Plenty of AI looks good in a demo. The harder part is building something people can rely on. A few habits guide how I do that.
The narrowest context, the fewest tools, the least data. A part of the system that can't reach something can't leak it or be talked into exposing it.
If a result changes from one run to the next, nobody can rely on it. I build pipelines that repeat exactly and point back to their sources, so results can be checked months later.
Audit logs, scoped access and a record of how each answer was reached. If you can't explain why the system said something, you can't stand behind it.
A prototype is a starting point. The work counts when it's running for real users, under real load.
Where I can help
Designing and shipping production AI end to end — agents, retrieval, guardrails, and self-hosted models where an API can't go.
Helping teams make the early decisions that are expensive to undo — before a lot of code depends on them.
Testing software, including AI features, with tooling that reasons about the specific app and proves each finding.
Contact
I'm open to AI engineering, architecture and training work. Tell me what you're building and what the constraints are — that's usually enough to start.