AI platform engineering
Engineering on Walma's distributed AI platform — a multi-tenant enterprise system that lets organizations chat against internal knowledge bases and run AI-powered document pipelines. Contributed across four interconnected codebases: Noda (Nuxt chat UI + RAG search and tool execution), Noda Workers (Azure Functions for ingestion and scheduling), Ocle (visual no-code pipeline builder), and Ocle Workers (19-worker serverless processing engine).
- Invoice Reviewer (Fakturagranskare) — rebuilt a municipal contract auditing workflow from 1 correct / 7 warnings / 3 inconclusive to 11/11 reliable outputs, via a decision-tree prompt with self-check stage, a deterministic extraction pass (temperature 0, fixed seed), retry logic, tool-call caps, and post-processing rules
- Platform architecture — established and applied the Ocle / Noda separation (stateless reusable document processing vs. customer-specific data, config, and scheduling), connected through a result/handler contract
- RAG ingestion pipeline — full path from Ocle Workers document parsing → webhook events → embedding generation and Azure AI Search indexing → hybrid search at chat time. Embedding migration from ada-002 to text-embedding-3-small and parent-child chunking for an RCO dataset
- Infrastructure consolidation — migrated Ocle Workers onto Noda's shared Azure AI Foundry hub, removing per-environment hub provisioning
- Tenant outcomes include a security-sector knowledge base (−65% support lookup time) and an invoice auditing workflow (−92% manual oversight)