Meharban Singh

Posts

• 2026-09-01 I Built an AI Software Factory. Here Are the Failure Modes Nobody Puts in the Demo. How a real multi-agent build system fails: stale state, tests that can't turn red, reviewers agreeing without new evidence, and what fixed it.
• 2026-09-01 My AI Agents Passed Every Test. The Software Was Still Wrong. Four failures from building a risk platform with AI agents: a UAT that never committed, a mutation nothing tested, a caller nobody wired, a metric that lied.
• 2026-09-01 'Done' Is a Claim, Not a Fact An installer shipped green with two unsubstituted placeholders. The fix that closed it required a different verifier, a RED/GREEN probe, and a database that refuses self-certified closes.
• 2026-09-01 The Writer Never Approves An orchestrator wrote 2,968 lines directly, then reviewed its own work. A bug in its own checker survived that review. What the role lock that followed actually changed.
• 2026-09-01 The Day I Rotated 15 Credentials A names-only secrets listing printed real values into a transcript, a redaction attempt made it worse, and the rule written afterward was broken again six days later.
• 2026-09-01 Shipping in a Domain I Knew Nothing About Building Astro Rattan meant shipping Vedic astrology without knowing the subject. The fix: take the model out of the answer path and let an expert verify.
• 2026-09-01 I Don't Prompt AI. I Run a Factory. A blueprint reviewed by a second model, hooks that block instead of advise, gates that read evidence not summaries, and a sentinel watching production — measured against 893 tracked issue rows.
• 2026-08-26 An Instruction Is Not a Guarantee Why a spec isn't enough once an agent does real implementation — and the process (skills, hooks, gates, independent review) I built around Claude Code instead.

Each post takes one real incident from the issue ledger — the failure, the root cause, and the rule or hook it became.