built against a public job description
DuvoAI — Enterprise-Retail FDE
An evaluation and tooling suite built against the DuvoAI Enterprise-Retail FDE role's public JD and Duvo's published materials: a deterministic agent-eval pack modeling the buyer-side retail/CPG supply-chain-ops agent the role describes, three companion packs, an MCP connector kit, and two small parsing libraries — one synthetic corpus behind all of it. Shaped around Duvo's own delivery motion — map the real process, automate only the approved version, govern every execution, keep the deployment re-runnable — rather than as a general portfolio piece.
SKU-###, PO-###,
SUP-###) produce every case.
The build, in five parts
Process intelligence
The process_mapping eval pack as the counterpart to a mapping engine — proving a reconstructed workflow before anything downstream trusts it.
Automations
The ten retail_ops tools, the MCP connector kit, and the browser-automation reliability gates — governed execution end to end.
Customer view
What an ops team sees once it's live: the supplier scorecard, the honest-ROI report, the exception-spec board they own.
Evaluation
Golden sets computed from fixtures, deterministic hard gates, keyless replay CI, and scorecard regression diffing.
Deployment
The per-customer scaffolder and the exception-spec DSL as the approval-gate configuration surface for an embed.
The operator behind it
Repos + writeup
agentic-eval-harness
The eval engine and its eleven domain packs — four of them new here, retail_ops and its three companions.
retail-ops-mcp-kit
Four FastMCP servers over the shared corpus, staged writes behind a human-approval token.
retail-ops-edi
Stdlib X12 850/856/810 parser, normalized to the pack's PO/ASN/Invoice shape.
retail-ops-uom
Pack-hierarchy UOM normalizer — returns an explicit Ambiguous rather than a silent wrong number.
Reproducible, keyless agent evals
The record/replay seam and the keyless-CI mechanism the packs above run on.