All projects
Case study · 2024 — Present
AgentOps — Production Agent Observability
Decision trees, token burn, hallucination signals
Independent builder · Independent
Personal product
RustClickHouseNext.jsReact FlowWebSockets
The problem
The problem
Agents fail semantically. Logs are huge. Teams cannot debug non-linear graphs with APM built for request/response.
Outcomes
- High-throughput log path on ClickHouse
- Real-time DAG visualizer for branching/looping agent runs
Approach
Ingest verbose monologue logs without drowning Postgres.
Stream graph updates over WebSockets into a React Flow canvas.
Key tradeoffs
Storage
Chose: ClickHouse for agent logs
Rejected: SQL OLTP only
Throughput and analytical queries dominate.
What was hard
- Visualizing loops without becoming spaghetti
What I would change
- Open-source core when redaction is clean
Artifacts
