Writeups
2026
- The Judge Problem: Why LLM-as-Judge Silently FailsPart 3 of the SeekerOS series — three judge failures that each broke the eval in a different way, and what to check before you trust a judge.
- Expensive Models Lost, Reasoning Models Lost: Benchmarking 16 LLMs for a Four-Step Pipeline16 LLMs, 48 test cases, four pipeline stages. The expensive models lost, the reasoning models lost, and three times a summary document carried a number the raw data contradicted.
- From Chat Agent to Pipeline: Building SeekerOSThe first post in a series on running LLMs in systems that have to be correct.
- The Browser Was the DashboardBuilding a local-first Chrome extension to replace the ritual of checking AI balances and usage limits across a pile of authenticated tabs.
- Zero to Cloud Run in a Day — Then the Bill, and the API That Said NoPorting a telemetry platform from Azure to GCP/Go with zero prior GCP experience, hunting down an API that silently blocks cloud egress, and cutting a $45/mo architecture to near-zero in one commit.
- The Pipeline That Lied: A Zero-Trust SQL SagaRemoving a SQL admin password kicked off a week of silent CI failures, an uncomputable binary SID, and a database with no name — every dead end preserved.