← All work

Agentic Decision-Intelligence Platform

Private commercial engagement

A commercial agentic decision-support platform I built as sole developer, from canonical data model to interface.

  • 2,807 graded checks
  • 151 browser assertions
  • Sole developer

Stack

  • Python
  • React + TypeScript
  • REST APIs
  • Canonical data model
  • Multi-agent workflows
  • Graded acceptance suite

A commercial agentic decision-intelligence platform, built as the sole developer across the whole stack: Python services and domain logic, a React and TypeScript interface, the APIs between them, and a canonical data model that everything else resolved against. Requirements moved throughout, so the architecture had to absorb change without the data model drifting underneath it.

The reasoning layer is source-grounded by design. Values carry their provenance and their uncertainty all the way to the interface instead of being flattened into a single confident answer, multi-agent workflows do the work of assembling and cross-checking, and a human approves anything consequential. The system is built to be questioned: a user can always ask where a number came from and get a real answer.

Release was gated on evidence, not on a demo. Acceptance ran as 22 independently graded families over a 2,807-check graded suite, plus 151 end-to-end browser assertions against the running application, and a longitudinal verification pass of 750 record-field checks that confirmed no factual value regressed and no provenance was lost between releases.

This was client work under a commercial engagement, and it is not my IP. The client, their data, the schemas, and the implementation stay confidential; what is described here is the engineering I owned.

A confidential commercial engagement, so the client, their data, and the codebase stay private. The architecture, the evaluation design, and the decisions behind them I'm glad to walk through on request.