Linaria

Auto-generated changelog categorized by Features, Bug Fixes, and Chores with real commit hashes and export options.
Full dashboard: daily standup, Team Velocity chart, Recent Commits, Team Activity with contributors, and Issue Board.
Self-Improvement Loop: LLM-as-a-Judge scores every response. Overall 6.8 across Helpfulness, Accuracy, and Tool Usage.

Inspiration

What it does

Inspiration

Small dev teams lose hours every week reconstructing context — what changed, who's blocked, what ships next. We wanted an agent that already knows the answer before you ask.

What it does

Linaria is an AI executive memory agent for small dev teams. It connects to GitHub and reasons over commits, PRs, and issues in real time. Ask it anything about your project: "What changed this week?", "Who's blocked?", "Generate the standup for today." It also runs a self-improvement loop — every response is scored by an LLM-as-a-Judge and the lessons are stored via Arize Phoenix so Linaria gets better between sessions.

How we built it

Agent: Google ADK + Gemini 2.5 Flash for orchestration and reasoning
Observability: Arize Phoenix for tracing + Phoenix MCP Server so the agent can read its own eval scores
Self-improvement: LLM-as-a-Judge evaluates every response; low-score lessons are saved to improvements.json and injected into the next session
Backend: FastAPI on Cloud Run with 15+ endpoints
Frontend: React + TypeScript + Tailwind + Vite on Vercel
GitHub integration: PyGithub for real-time repo data

Challenges we ran into

Getting the self-improvement loop to actually close was the hardest part — having the agent read its own Phoenix traces via MCP, extract lessons, and apply them in the next session required careful prompt engineering and span annotation design. We also had to isolate the LLM evaluator from the critical response path so a slow judge never blocks the user from getting their answer.

What we learned

LLM observability isn't just for debugging — when the agent can read its own traces, it becomes a feedback mechanism. Arize Phoenix + MCP turned a passive monitoring tool into an active part of the agent's reasoning loop.