DAC Platform
Agentic knowledge-graph RAG · AIBC
AIBC's evolving agent platform: DCS (smart RAG) → DAC (knowledge-graph + vector RAG) → DACv2 (autonomous, Medusa-powered). The aim is to turn unstructured documents into something an agent can reason and act over.
Problem
Companies sit on piles of unstructured documents — contracts, reports, scans — that plain vector search can retrieve but can't reason over. Similarity search finds passages; it doesn't understand how entities across a corpus relate, and it certainly can't act on what it finds.
Approach
I led the platform through three deliberate generations rather than one big bet. DCS proved smart RAG: OCR, chunking, hybrid retrieval. DAC added a knowledge graph beside the vector store, so retrieval could traverse entity relationships instead of only matching embeddings.
DACv2 made it agentic: autonomous multi-agent workflows running on Medusa, the reusable runtime I built underneath it — MCP-native tool access, multi-provider LLM adapters, and human-in-the-loop governance gates for anything irreversible.
Outcomes
- Three shipped generations of the platform, each in production use inside AIBC's pipeline.
- The architecture split (platform vs. reusable runtime) let Medusa outlive the product it was built for — it now powers other agent work, including the thinking behind Moirai Flow.