Jenish Kothari
About
Software engineer who likes hard problems and small models that overperform.
8–4: System design @ Adobe. Authorization infra and asset approval workflows that stay fast and correct at enterprise scale.
4–10: Post-training LLMs + agentic AI. Recently fine-tuned a 7B model to beat Grok-4 on Text-to-SQL for $25.
Building
VibeSec, a security benchmark for AI-generated code. 1,000 tasks, each a stable URL carrying the prompt, the generated app, an exploit that proves the bug, and a reference patch. Frontier models are scored on whether the code they write survives it.
Interests
- Post-training. SFT, reasoning distillation, DPO, GRPO. Evaluation that executes the output instead of string-matching it, and local models above all: a 7B that fits on one GPU and matches a frontier model is worth more to most teams than an API call.
- Competitive programming. Aiming for the ICPC World Finals.
- Badminton, table tennis, pickleball.
Achievements
- Winner · Confluent AI Day 2025 · $2,000 (SecureStream AI)
- Winner · AI Builder Day (MadeThis) · $1,000 (SmartGTM)
- Winner · Build with Gemma (Kaggle × JustBuild) · $250
- Guardian @ LeetCode (1,897 solved) · Expert @ Codeforces (max 1618) · ICPC North America Regionalist
Competitive programming
Writing
| Post-training | The starting line: a 7B against the frontier · 6 Nov 2026 Part 2 of Post-Training 101. Qwen2.5-Coder-7B with no training, next to three much larger models, with and without BIRD's evidence hint, and how far each score can be trusted. |
| Post-training | Understanding the data and the benchmark · 25 Oct 2026 Part 1 of Post-Training 101. Why I am post-training a 7B model to write SQL, what I want to learn along the way, and a first look at BIRD, the benchmark it will be graded on. |
| Interviews | Two rounds at Stripe, and one part short · 6 Oct 2026 A cold application to Stripe's new grad Software Engineer role in November 2025. A hard four-part online assessment, then an interview in the same format where I finished three parts of four. The biggest mistake was the language I chose. |
| Interviews | Two rounds at Meta, and a rejection within the hour · 5 Oct 2026 A cold application to Meta's 2026 Production Engineer University Grad role. Two screen rounds on systems fundamentals and coding, a rejection within an hour of finishing, and then silence. |
| Interviews | Four rounds at ServiceNow, and results that came too late · 3 Oct 2026 The ServiceNow SWE internship loop for summer 2025. A recruiter screen, then three 30 minute interviews in one afternoon covering systems design, two LeetCode mediums and containers. The results were slow to come back, and by then I had already said yes to Adobe. |
Elsewhere
Start here · Projects · About
GitHub · Hugging Face · X · LinkedIn · kotharijenish2001@gmail.com