01
Sxored
Mar 2024 to Present
A hybrid pipeline of deterministic parsing and AI that turns financial documents into decisions credit analysts can act on.
Problem Credit analysts were reading financial documents by hand: no consistent format, no shortcuts, and a decision that has real money behind it. Pure AI extraction was not an option either, because a hallucinated number in a credit file is worse than no number at all.
What I owned The architecture and the hybrid extraction pipeline: deterministic parsing where accuracy is non negotiable, LLM and VLM judgement where structure runs out, every field traceable to the page it came from. Event driven Go services orchestrated with Temporal on Kubernetes, a React back office, the infrastructure bill, the engineering team, and the ISO 27001 controls banks audit.
Where it stands In production with banks and lenders today, alongside a quick appraisal service that gives appraisers a fast, defensible price for a property being sold or pledged. Still my week, still my architecture.
Stack
Go microservicesReactTemporalRabbitMQKubernetesPostgreSQLLLM / VLM