OpenTriage: A Read-Only Triage Assistant for Maintainers, and Why a No-LLM Baseline Beat Gemma on Precision

This is a story about OpenTriage, a local-first triage assistant designed for open-source maintainers. It helps review issues and pull requests by suggesting labels, flagging duplicates, identifying low-effort PRs, and drafting polite replies. OpenTriage runs on a laptop using small open-weight models like Gemma and never labels, comments, or closes anything on GitHub. The human always decides.

The story highlights the challenge faced by maintainers during Hacktoberfest when the flood of issues and PRs creates duplicate reports and trivial changes. OpenTriage was built to alleviate this burden for maintainers and open-source contributors.

The development stack used for OpenTriage includes Next.js, TypeScript, Tailwind, Zod for schema validation, and Ollama for local AI model execution. The model of choice is Gemma (gemma3:1b via Ollama) to re-rank labels and draft replies. The system utilizes local embeddings and BM25 lexical scores for duplicate detection, but it can function without embeddings.

The live demo of OpenTriage can be accessed at https://opentriage-0xzg.onrender.com/. For the hackathon, the demo was tested on the expressjs/express and colinhacks/zod repositories. The full implementation is available on GitHub at pkprajapati7402/OpenTriage.

The story emphasizes the importance of privacy and cost-effectiveness for open-source maintainers. With OpenTriage, sensitive data stays on the maintainer's machine, and running the assistant locally is free. The project was built for the DEV Hacktoberfest 2026 Weekend Challenge: Build for a Friend.