Enterprise Document AI Assistant / the ingestion backbone

From full reprocessing to incremental sync

I owned technical direction for the document-processing service — the ingestion backbone of an enterprise AI assistant platform spanning five microservices. The hardest problems were rarely the LLM: they were keeping a large SharePoint corpus in sync, respecting permissions at scale, and staying reliable under Microsoft Graph throttling.

Redesigning the sync architecture around Graph delta tokens cut the workload from full-corpus reprocessing to incremental updates only. When platform-wide throttling failures hit, I architected a dedicated permission-resync worker with backoff and retry that restored reliability for large-scale enterprise syncs.

What changed

• MCP-based document access controls, enabling a new Personal AI Assistant capability across the platform.
• Multi-modal LLM features — image generation, translation, and a Q&A agent — delivered across two backend services.
• Multi-database usage and cost reporting that gives leadership real visibility into how the platform is used.