Student builds a local coding agent on molab and beats two 7B coder baselines
S_Conradi · x · 2026-07-29
- A final-year CS student built Chitti, a local coding agent for security work, because existing assistants were too cloud-dependent and often refused security-adjacent tasks.
- He handled the whole pipeline himself on molab: dataset curation, fine-tuning, benchmarking, and deployment, all from a notebook.
- The project emphasizes privacy for proprietary codebases and a more permissive workflow for tasks like auth flows, validation, and network code.
- The chart shows Chitti outperforming DeepSeek-Coder-7B and Qwen2.5-Coder-7B on HumanEval+ (76.83%), LiveCodeBench (23.0%), and MultiPL-E (67.49%), with a 4.7 GB 4-bit quantized model size.
More from coding & agent
- YC Startup School says coding agents still need human context in production incidents — ycombinator · 2026-07-29
- Hackathon team uses OpenAI Codex to drive a robot arm from natural language — davidfromkansas · 2026-07-29
- A 4-loop framework for production AI agents focuses on task, verification, and improvement — MaryamMiradi · 2026-07-29
- GitHub Copilot App is adding a pull request view for model-based review — mariorod1 · 2026-07-29
- Developer patches Codex CLI to render LaTeX directly in the terminal — BLUECOW009 · 2026-07-29
- LangChain Guide: How to Provision Secure Execution Environments for AI Agents — LangChain · 2026-07-29