IntelligentOldMan AI operations platform for a trading-card business
- Built an event-driven pipeline that turns a Microsoft 365 mailbox into queryable business data: Graph change notifications flow through Event Grid and Service Bus, then each message is deduplicated, classified, grounded, chunked, and embedded into pgvector.
- Implemented RAG question answering with
QuestionAnswerAdvisor so the agent answers only from retrieved email context, and escalates high-signal supplier inquiries to a human.
- Added an intent classifier with a grounding evaluator and an operator correction API, so mistakes feed back into the system instead of disappearing.
- Wrote Dokimos eval suites (intent accuracy, grounding judge, RAG answer quality) that run as a separate Gradle task, keeping LLM regressions out of main.
- Shipped a card identification API (image and OCR text to catalog lookup) and exposed agent helpers as
@Tool methods for MCP.
- Ran the infrastructure as Terraform stacks on Azure, including Graph subscription lifecycle management, delta-sync reconciliation, and OpenAI capacity and 429 throttling controls.
sdvosb-brainDaily SAM.gov ingest that summarizes and scores SDVOSB contract notices against a capability statement, using local Ollama models and pgvector.
Voice Agent referencePhone agent on the OpenAI Realtime API with RAG over pgvector and backend tool calls, with Spring AI as the brain and a small Node bridge for audio.
RAG JudgeSpec-driven design for a pairwise judge that compares RAG answers, planned as 18 vertical-slice tasks.