I build Apify actors — small, focused tools that turn raw scraped data into something a real workflow can use. Lead enrichment, dataset transformations, AI-powered field derivation, automation glue between scrapers and downstream systems.

• Ship narrow, ship measured. Every actor I publish targets one specific pain point and publishes precision / recall / FP-rate numbers against a human-labeled eval set. If I can't prove the quality, I can't justify the price.
• Null is a feature. When my actor can't confidently produce a result, it returns null — and doesn't bill you for that record. Hallucinating fake data to inflate recall is a trust-debt I won't take on.
• Open methodology. Every actor's GitHub repo includes the labeling guide, the comparator code, and the eval results. Want to audit my published precision numbers? You can.
• Honest pricing. Pay-per-event aligned with delivered value, not per-second-of-compute. You pay for results, not retries.

• PRD-driven: every actor starts with a locked spec. Scope creep is the enemy of shipping.
• Eval-gated: no public release without a labeled eval set and a published numbers table.
• Provider-portable: LLM-backed actors run on AWS Bedrock with model fallbacks documented and env-flippable. No vendor lock-in.
• Boring tech: TypeScript strict, Apify SDK v3, Zod at boundaries, Vitest. Predictable beats clever.

TypeScript (strict) · Apify SDK v3 · CheerioCrawler · AWS Bedrock (Converse API) · Zod · Vitest

• Apify Store — drop an issue on any actor page

I take hallucination reports and quality-regression bugs seriously. Every false-positive someone shows me gets added to the next iteration's eval set.