Most of my work is for organisations where the data is sensitive and the output gets scrutinised: health plans, legal practices, a government agency. In that setting, a clever prompt isn't enough. What matters is where the data goes, whether an answer can be reproduced, and whether someone can check how it was reached.
So I spend a lot of time below the prompt layer — isolating sensitive data, making pipelines deterministic, and fine-tuning, compressing and self-hosting models when an API can't meet the cost, latency or data-residency requirement. I also build AI tooling for security testing, which has taught me a lot about where systems break.
Alongside the engineering, I teach. Explaining this material to a room of engineers is the quickest way I know to find the gaps in my own understanding.
Model work
Three things push you below the prompt layer: cost that stops being small once you're at volume, latency you can't fix from outside, and data that isn't allowed to leave the building. A hosted API doesn't solve any of them.
So you host the model yourself — and everything the API was quietly handling becomes your job: how much VRAM it needs, how fast it runs under concurrent load, and how much capability you give up to make it fit. It's less visible than prompt design, but it's where a lot of the real engineering happens.
Teaching a model the domain without training from scratch.
Making it fit without making it useless.
Throughput is a design decision, not a number from a benchmark page.
The groundwork that still shapes most of the decisions above.
Tools I use
In short
Common questions, answered plainly.
Ajith Thaduri is an AI engineer and technical consultant based in Hyderabad, India. He designs and builds production AI systems — agentic architectures, retrieval pipelines and self-hosted language models — for healthcare, legal and government organisations where the data is regulated and the output is closely checked. He also builds AI-driven security tooling and teaches engineering teams.
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.