someshbhardwaj.dev rebuild
· LIVE · platform
Rebuilt this site as a supervised multi-agent build, with the verification gates written before any of the code.
Problem
The site contradicted its own positioning: it read as an AI-agent portfolio while the work underneath was systems and operations. The layer beneath was worse: no shared style abstraction across 603 inline class attributes, no router so every URL rendered the home page, and structured data publishing metrics that existed nowhere in the repo. There was no CI, and seven tests.
Intervention
Ran it as an engagement rather than a rewrite. Gates first: typecheck, lint, CI, and tests targeting the couplings that fail silently. Then a frozen written brief: design tokens, banned patterns, tone, and file ownership, that every agent read before writing a line, with disjoint file sets so four could work in parallel without collision. Built a console that reads the agents' own transcripts to show what each actually touched, rather than what its instructions said it would. Every claim the agents produced was verified against the source before it shipped.
Result
- Tests 7 \u2192 79, including guards for failures that produce no error
- One indexable page \u2192 35, each record entry its own URL
- 32 record entries, every date carrying its provenance
- The gates caught real errors: a runtime crash, a five-month date mistake, and an overstated integration claim I had written myself
- An agent declined to invent a response-time figure because the brief forbade unsourced metrics
Stack: React 19, Vite, Tailwind v4, TypeScript, Vitest, Netlify, Multi-agent orchestration