About
Hi, I'm Shrikar.
I'm an AI-native engineer based in the Bay Area. I build production AI systems and fine-tune small models for the narrow tasks where accuracy, latency, or cost matter, and I like helping teams level up the way they actually operate AI.
What I work on
Structured-extraction and narrow-task problems where the frontier-LLM API is too slow, too expensive, or too unreliable. The pattern I keep coming back to: a fine-tuned small model that beats the general API on one specific task, typically at a fraction of the per-document cost and latency. I also spend a lot of time on agent harnesses and eval pipelines, the unglamorous plumbing that lets a team run AI systems with the same rigor they bring to the rest of their stack.
A recent example: I fine-tuned a Llama 3.1 8B model that extracts structured JSON from Bills of Lading at 99.6% field accuracy and 89 ms per document. That beat Claude Sonnet 4.5 (92.4%, 3,981 ms) and Gemini 3.1 Flash Lite (91.9%, 1,699 ms) on the same test set, at roughly 1/18th the cost per document. Full case study.
How I approach it
-
Diagnose first
Run the real data through the frontier model, score every field, and find the specific failure modes a smaller model could fix. If a general model already hits 99% on the task, that is the answer. Fine-tuning is theater when it isn't necessary.
-
Data, then training
Build a training corpus with realistic distribution variation (layouts, OCR noise, missing fields). Train, evaluate, iterate against two adversarial test splits, never just an in-distribution holdout.
-
Make it reproducible
An inference path that actually runs in production, plus a monitoring and eval harness, drift detection, and a documented runbook. The result should hold up without me in the room.
Where fine-tuning is the wrong tool
- General-purpose tasks (summarization, open-ended QA). Use the API.
- Low volumes. Prompting usually beats fine-tuning on total cost.
- Problems without ground-truth eval data. Fix the data first.
Let's talk
I mostly write here to think out loud about this work. If you're wrestling with a structured-extraction problem, an eval setup, or anything AI-native and want to brainstorm or compare notes, I'd genuinely enjoy that. No agenda.