Three model calls, and none of them reads a price

Description and use

Nightly repricing runs in 24 min, costs <$2, and delivers accurate, time‑stamped prices per mileage tier with transparent formulas. The system scales to zero, uses minimal models, and avoids silent failures.

Problem

Automated nightly repricing of car offers was unreliable because advertised prices omitted mileage and validity conditions, leading to misleading comparisons.

Outcome

Nightly repricing runs in 24 min, costs <$2, and delivers accurate, time‑stamped prices per mileage tier with transparent formulas. The system scales to zero, uses minimal models, and avoids silent failures.

Workflow

• Schedule nightly Cloud Run Job via Cloud Scheduler
• Read source registry from Firestore and apply refresh policy
• Fan work to Pub/Sub; failed tasks go to dead‑letter topic
• Scrape offer pages, store HTML snapshots in Cloud Storage with SHA‑256
• Run three model calls: blind verification, deduplication, terms harvest
• Perform blind verification by sending stored HTML to Gemini 3.7 Flash without parser output
• Deduplicate offers using gemini‑embedding‑001 and deterministic rules
• Harvest terms‑and‑conditions from PDFs via Cloud DLP and Google ADK
• Swap new data transactionally, keeping yesterday’s data live until ready
• Publish pricing formula on site, versioned with changelog

Evidence

The run finishes with a single transactional swap: until the whole set is ready, yesterday's data stays live.

Source

The whole night, models included, stays under $2.

Source