About

I'm Jenish Kothari. Software Engineer II at Adobe in Salt Lake City, and an MS CS graduate of Northeastern University (2026). Before that, BS Computer Science at Nirma University.

My work is post-training: SFT, reasoning distillation, DPO, and GRPO. At Adobe I distilled a production LLM down to a 3B parameter model, cutting inference cost 10× while holding 72% of GPT-4 benchmarks. On my own time I do the same thing for text-to-SQL, evaluated by executing every generated query against the real database and comparing result sets.

The thread running through all of it is local. A 7B that fits on one GPU and matches a 1.6T frontier model is worth more to most teams than an API call, so that's what I optimize for. See Projects for the numbers.

Experience

Recent platform work: migrated access control for 5,000+ approval templates from MongoDB to a Zanzibar-style AuthZ service with a phased backfill and zero-downtime cutover, plus an automated reconciliation audit that caught orphaned templates from partial-failure edge cases before they reached customers.

Published at IEEE ICAD 2026. Won the Build with Gemma hackathon (Kaggle × JustBuild), the Confluent Kafka hackathon, and Best Project at the Harvard AI Hackathon. Mostly Python and PyTorch, with whatever else the problem needs.

Contact

Most responsive on X DMs. Email: kotharijenish2001@gmail.com. Also on LinkedIn, GitHub, Hugging Face, and Kaggle.