Blogs
Notes on systems, observability, and the things I break at work.
-
Multi-agent debate for diagnostics
Jul 2026A two-phase diagnostic system: extract and tier evidence first, then run a specialist panel that debates hypotheses with grounded tool use — architecture patterns that transfer beyond the original fermentation use case.
-
Building a cost-effective CI pipeline at Rocketium
Jul 2026How we built an in-house CI service to handle dependent PRs, approvals, merge conflicts, and readiness checks without paying per-seat SaaS pricing — scanning repos, connected components, and merging ready units to patch.
-
Druid segments, partitioning, and the hot key that broke my assumptions
Jul 2026What a Druid segment actually is, how primary and secondary partitioning work, the difference between hash and range partitioning, and how one very large tenant quietly produced segments many times larger than the configured limit, with the ingestion tasks reporting success the whole time.