Instance logs: ClickHouse, Sentry, Kafka, Prometheus
Every VM line is stored in ClickHouse. The studio shows 20 KB. Errors go to Sentry. Kafka is an optional buffer; Prometheus is metrics only.
→ ReadGraphai › Jobs › Blog
Jobs blog
How the apply fleet is built, what it costs, and where job listings come from. Public notes only — the jobs product is not released. Company pages appear here only after Graphai has added research from Wikidata, Vertex, Cala, or Tavily.
Company profiles with extra research
How job listings are extracted
Career HTML, ATS JSON, and job-alert mail become ClickHouse rows. PyTorch clusters duplicates; URL scores prefer Greenhouse over Easy Apply; skills go past the 20-term catalog; demand is forecast per title.
→ ReadAWS and GCP subsystems for apply and crawl
Every Amazon and Google service that is live, planned, or off — EC2 Spot, SSM, CloudWatch, CloudTrail, S3, Bedrock, Compute Engine, Cloud Run, Vertex Gemini — each with its official docs link.
→ ReadApply fleet data architecture
Day-0 ingest from job-alert mail, AWS €100 and GCP €200 caps, Vertex Gemini vision on failing screens, and why LinkedIn Easy Apply is out of scope.
→ Read