Anyone can call an LLM on one row. The hard part is calling it on twelve million domains, across 77 country markets at a cost and latency that lets you actually run a SaaS business on top of the results. This session is a deep technical walkthrough of how DomainScope built a production enrichment pipeline that uses LLMs as a classification and extraction engine, turning raw, multilingual web data into structured market positioning, ecosystem maps, and concentration signals.

We will cover: the Go architecture that fans out millions of jobs across a fleet of 150 microservices; why we only use LLMs for tasks they excel at, and then use LLMs in a loop to iteratively build the ultimate CPU-only microservices for everything else, cutting recurring inference costs to near zero; how a shared fleet-wide cache allows a single scraped page to be reused across all 150 services instead of re-fetching or re-processing per job; the prompt-evaluation loop that lets us safely swap models without regressing 12M existing rows; and the credit-metering API design that exposes this externally without leaking the cost structure.
Attendees leave with a concrete pattern for designing AI-powered data APIs where the LLM is an internal tool, not the product, including the failure modes, retry topology, observability (Prometheus + Grafana), and idempotency guarantees that make the pipeline survive provider outages and rate limits in production.

- Main Site: https://domainscope.scrapetheworld.org/
- Map View: https://domainscope.scrapetheworld.org/map
- Country Index (e.g., US): https://domainscope.scrapetheworld.org/market-index?country=United+States
- City Index (e.g., San Francisco): https://domainscope.scrapetheworld.org/market-city/San%20Francisco