Full-stack AI engineer  /  game builder

The model proposes.The code decides.

I build production AI systems where an LLM runs only where natural-language judgment is unavoidable. Everything that can be deterministic is, and the boundary between the two is tested.

See the projects Work history Email
scroll
the shape of the work

Every system I build looks like this

A deterministic spine does the work that can be specified. The model is one stage in the middle: it reads intent and proposes, then plain code enforces the parts that carry consequences, before anything is said back.

request validate dedupe enrich LLM proposes intent · reasoning confidence enforce plain code amounts · rules routing · ~60 tests response
the same idea, off the clock

A cricket RPG built on the same rule

I am building a career-mode cricket game for mobile. A deterministic match engine resolves every delivery from footwork, shot and timing. The language model never touches the outcome. It sits on top as a paid commentary layer, and free players make zero model calls.

4
engine decides
leg glance · 4 runs · timing 0.86 · edge 0.0
llm narrates
"On the pads, and he glances it fine off the hip. Beats the keeper, four."
Unity 6 URP  ·  Python + TypeScript engine  ·  Blender character pipeline

Start here

Projects
The support bot tuned against a grader, Sreshtha, the OpenAI Parameter Golf contribution, and more.
Work
Flipkart, and seven years at Swiggy building decision engines and an in-house AI QA engine.
Profile
Skills, environment, education, and what I build outside the day job.