← All case studies

AI Engineer

MarketBrew Oracle Agent

An eval-first agent on top of the MarketBrew knowledge base, with answers cited back to the atoms.

Private system, no public URL

There is no public URL for this one. The write-up below covers the architecture and the decisions behind it, and we can walk you through it on a call.

The agent layer built on top of the MarketBrew knowledge base (see the Reference Manual project). An eval-first agentic scaffold on PostgreSQL: contracts and fixtures are defined before generation, and every recommendation is evidence-cited back to the underlying knowledge atoms. The scaffold pattern transfers across client domains.

MarketBrew Oracle Agent: a recommendation with each claim traced back to a knowledge atom

The problem

Most LLM-backed recommendation tools generate first and validate later. When the generation is wrong, the team is left chasing why. The eval-first pattern flips that order: define the contract, build the fixtures, generate against the contract. Outputs are checkable from the moment they exist.

What we built

A PostgreSQL-backed agent layered on the MarketBrew knowledge atoms, with contracts and fixtures defined before any generation happens. Every recommendation cites the supporting atom, so a reviewer can trace the path from output back to evidence. The eval-first scaffold is the deliverable: a foundation that is checkable from the first output rather than debugged after the fact.

What was delivered

Outcomes

Services: PostgreSQL, Anthropic API, Node.js, Eval-first design