01 / Source
Hitarth Desai
I build the layer that makes LLMs reliable in production.
Currently at InfraDock AI, designing LLM-powered document processing pipelines. Previously 2× AI/ML Lead at GDSC Surat. Computer Engineering, GTU 2026.
02 / Extract
What I work on
I care about the gap between a model that demos well and a system that survives real traffic. Most of my work lives in that gap — turning probabilistic model output into something a downstream service can trust, then making it fast and observable enough to ship.
- Reliable extraction — schema-constrained, validated, deterministic structured output from LLMs instead of "parse the JSON and pray."
- Pipeline infrastructure — async, batched, observable document-processing flows.
- Open source — I publish the tools I needed and couldn’t find.
03 / Structure
Selected work
A multi-agent prompt-optimization workbench for systematically testing and improving prompts against eval harnesses. In active development.
AEO diagnostic tool (Gemini + Firecrawl) that scores how discoverable a page is to AI answer engines.
My first venture: a dataset-intelligence platform for inspecting, profiling, and cleaning ML training data.
04 / Stack
Tools I reach for
- Languages
- PythonTypeScriptRust (learning)
- LLM / Data
- structured outputsconstrained decodingmsgspecLangChainvLLM
- Infra
- DockerKubernetesTerraformKafkamicroservices
- MLOps
- async pipelinesbatchingeval harnessesobservability
05 / Trace
Where I’ve worked
- Now
AI Systems Engineer · InfraDock AI
Designing LLM-powered document-processing pipelines — turning probabilistic model output into structured data downstream services can trust, fast and observable enough to run under real load.
- Previously
2× AI/ML Lead · GDSC Surat
Taught and built alongside a few hundred student engineers across two terms leading the AI/ML track.
- Earlier
Backend & systems work · Independent / contract
Shipped production systems ranging from event-driven backtesting infrastructure to microservice platforms.
- Education
B.E. Computer Engineering · Gujarat Technological University
Graduating 2026.
06 / Query
Common questions
What does Hitarth Desai do?
He is an AI Systems Engineer at InfraDock AI, building LLM data-extraction pipelines, deterministic structured output, and the MLOps infrastructure that keeps them reliable in production.
What open-source tools has he built?
confident-extract — a Python library for deterministic, schema-constrained extraction from LLMs, built on msgspec — is published on PyPI. promptcrucible, a multi-agent prompt-optimization workbench, is in active development.
What roles is he open to?
AI Systems and Founding Engineer roles at seed to Series A startups.
Where is he based?
Surat, Gujarat, India.
What is his background?
A Computer Engineering graduate (GTU, 2026) and 2× AI/ML Lead at GDSC Surat, where he taught and built alongside a few hundred student engineers.
What makes his work distinctive?
He focuses on the reliability layer between a model that demos well and a system that survives production traffic — extraction, validation, batching, and observability.
07 / Publish
Writing
Field notes on making LLMs reliable in production — structured outputs, pipeline infrastructure, and the unglamorous layer in between.
08 / Signal
Off the clock
International gold-medalist powerlifter. I play guitar and perform live — mostly metal and hard rock. The discipline of progressive overload turns out to map surprisingly well onto debugging distributed systems.
09 / Ship
Get in touch
Currently at InfraDock AI. Open to AI Systems / Founding Engineer roles at seed–Series A startups.