LLM-powered lead-sourcing engine: Amazon Bedrock
- Type
- Project
The lead-sourcing engine scanned over 1.3 million Nordic websites through Amazon Bedrock, served through FastAPI with Pydantic models validating and structuring every request and response passing through it.
The real work was making an LLM-orchestration pipeline reliable at that volume, not just getting it working once: I hardened it to run unattended against real throttling and transient failures rather than letting one bad response kill an entire batch, and built out result publishing so the analysis was genuinely usable once it finished.
Getting an AI pipeline from "works in a demo" to "runs unattended against 1.3 million real websites" is mostly this kind of operational hardening, and it's the part of AI engineering that doesn't show up in a model card.