Trust comes from structure
Separate what LLMs do best from what code does best — the design principle running through every Ghil Labs system. Click any node in the diagrams below to see what each component does.
Watch · Shorts
The design principle, in under a minute
Dual-track retrieval that separates private and public knowledge
Sensitive personal asset data stays local; shareable knowledge lives in the cloud. One embedding model (BGE-M3) powers both tracks, and an MCP server makes retrieval available to any AI client.
LLMs judge, code calculates, humans decide
DealDesk's underwriting pipeline. A coordinator directs subagents, all numbers come from a deterministic calculation tool, and anything below the confidence threshold escalates to a human.
Three rules applied to every system
Numbers are computed by code, not LLMs
DSCR, cap rates, tax simulations — every figure that must be verifiable is produced by deterministic functions. The LLM interprets and explains; it does not do arithmetic.
Private data and public knowledge are physically separated
Asset and tax data exist only on the local track (ChromaDB). Only shareable knowledge goes to the cloud (pgvector). Privacy is never traded for convenience.
When uncertain, hand it to a human
Below the confidence threshold, the system escalates instead of inventing an answer. AI is the verifier, not the decision-maker — final judgment always belongs to a person.