Build an enthusiast car search site that compiles vehicle history dossiers

Pacione, a database developer turned manager who no longer codes, wanted a site where buyers of used performance cars can see a fuller history of a specific vehicle. He prototyped with ClickHouse's built-in agent and Lovable, then moved everything to Cursor on the Pro+ plan, mostly in Auto mode. Cursor wrote the TanStack front end, the ClickHouse and Neon schemas, and ingestion pipelines that pull MarketCheck listings, NHTSA data, web search results through SerpAPI and auction records. It also wired up Better Auth, Cloudflare Turnstile, Resend email and vitest. Search understands chassis codes, factory colours, options and past maintenance. Grok, Sonnet and Gemini were used to review the code and architecture. Daily and weekly jobs run on his MacBook Air.

The site works and covers several makes, from BMW M and Porsche to Ferrari and McLaren. Each integration took minutes rather than the weeks he expected. Cursor preferred deterministic parsers and only fell back to an LLM, which kept costs down. The misses: no layered data model, slow searches from parsing large JSON blobs on every request (fixing it cut daily ClickHouse cost 30 to 50%), plain HTTP calls instead of a ClickHouse driver causing pipeline timeouts, near-duplicate pipeline code for every new make that had to be rewritten, and fixes that broke earlier searches. Running costs are about $800 a month for MarketCheck and $500 for ClickHouse.

Reviewer models did not catch the JSON or driver problems either. What helped was adding written contracts on code reuse, TDD rules, a test harness and architecture decision records. One car with nearly 600 historical listings used 12% of a month's data budget, so cap per-vehicle API calls early.