Start Here
I'm Jenish. Software engineer at Adobe, working on post-training. The work I care most about is making a small model that runs on one GPU do what people currently pay a frontier API to do, and proving it with evaluation that executes the output instead of string-matching it.
| Post-training | Text-to-SQL on a small model: execution-verified study on Qwen2.5-Coder-7B/14B, benchmarked on BIRD. A 7B ties DeepSeek V4-Pro (1.6T) at roughly 0.4% of the parameters. |
If you're learning DPO: DPO Text-to-SQL Lab: six notebooks from y = 3x + 2 up to
DPO on real preference pairs.
| |
| If you're learning GRPO: grpo-text-2-sql: the SQL executor is the reward function, so there's no reward model and no preference pairs to build. | |
| Local | qwen2.5-coder-7b-bird-cot: reasoning-distilled text-to-SQL model. GGUF build for local inference. |
| LocalSQL: ask a question, get SQL, run it live. | |
| Evals | Datasets: bird-cot-sft (CoT traces distilled from Qwen3-Coder-480B), spider-dpo-1040 (1,040 execution-verified preference pairs). |
| sql-agent-rl-env: the RL environment for SQL agents. | |
| Agents | package-quarantine: detonates dependencies in a honeypot sandbox before your coding agent installs them. |
| MedSignal is local-first clinical intelligence, running entirely on-device. | |
| Systems | Drift Sentinel: catches the schema changes that quietly degrade ML models without breaking any pipeline. |
| SplitMate: receipt photos to itemized group splits. 22+ users, $23K+ split across 182+ receipts. | |
| Research | Automatic Highlight Generation, IEEE ICAD 2026. paper repo. 82% accuracy, 88% recall on goal detection; condenses a 90-minute match in under 15 minutes. |
Everything else is on GitHub and Hugging Face. Most responsive contact: X DMs.