Document Intelligence over Microsoft Teams — from full re-scans to delta sync

20+ executives and team leads rely on this module weekly. Reports auto-ingest from Microsoft Teams via the Graph API; delta tokens replaced an ~18-minute full re-scan per cycle with an incremental change feed. A dual index — page-image embeddings (ColQwen2.5) fused with BM25 sparse text — resolves chart-heavy slides and exact keyword lookups in one search. Answers stream from a 13-tool agentic loop with resumable human-in-the-loop approvals.

Failure-Analysis Automation — from a searchable knowledge base to generated drafts

Built the failure-analysis report knowledge base. A two-pass vision-LLM pipeline segments inconsistent, multi-case decks (every author formats them differently) into individually embedded cases for cited, agentic Q&A. Then I flipped the problem: a generator drafts standardized reports straight from log data, building the charts, commenting on the data, and flagging likely causes. It's a draft, not a verdict: the engineer concludes, and every number is computed in code, never by the model. The drafts feed back into a cleaner knowledge base.

Governed Text-to-SQL Analytics — from spreadsheet-locked data to self-service

One document set is mostly spreadsheets, so instead of retrieval this module runs plain-language analytics via Text-to-SQL on DuckDB. Every generated query passes a defense-in-depth check — sqlglot AST validation, SELECT-only whitelists, forced row limits, read-only connections, and one self-correcting retry — and a governed semantic layer exposes only admin-verified tables to the model. Non-technical users can also compose personal agents from a whitelisted tool palette, no code required.

AWS Migration — from on-premise to AWS

Designed the migration with LG CNS — architecture and cost model — and I'm now executing it: migrating the production platform to AWS, making the cost and architecture calls that keep it viable.