W012026AI enablementTransportation, contract3 vendor teams

Outcome

40%

fewer senior architect escalations, after the support agent sustained 95%+ evaluated accuracy against a reference response set.

675operations migrated across 18 phases into an agent-ready context pipeline, zero hard deletes

8session applied AI curriculum built and taught to 12 product, project, and executive leaders

42percent of support tickets needing 3+ comment exchanges before the real request surfaced

The agent that reads tickets so architects don't have to

01 / Context

I own AI enablement across the whole organization — production agents, the skills program, and the context architecture underneath both — while concurrently managing three vendor development teams across Data, DevOps, and Development.

02 / Problem

A runtime override was making agent behavior changes with nobody reviewing them.

The org had agents running, but changes to their behavior weren’t source-controlled — there was no PR, no review, no audit trail. Separately, support intake was quietly broken: 42% of tickets needed three or more comment exchanges before anyone could tell what was actually being asked.

PR-GATED CONFIG FLOW — DIAGRAM

FIG 01 — Runtime override to reviewable PR.

03 / Approach

Fix the governance gap first, then build the agent, then prove it works before it ships.

01Convert the unsafe runtime override into a PR-controlled configuration, so every agent behavior change is reviewable before it ships.

02Build the Knowledge Curator to read resolved tickets, cluster recurring patterns, and maintain runbooks autonomously — the pattern-recognition work a human was doing by hand.

03Evaluate before releasing anything: build the support Rovo agent against Knowledge Curator runbooks in Rovo Studio, and withhold release until it sustains 95%+ accuracy on a reference response set.

04 / What I did

Designed, hardened, and shipped the Knowledge Curator agent end to end, then used it as the foundation for a second agent (support Rovo) evaluated against it. In parallel, restructured the Data team’s entire knowledge base into an agent-ready context pipeline — a ~675-operation migration across 18 phases, delivered on schedule with zero data loss despite a mid-effort platform API failure.

Also built and taught the org’s first applied AI curriculum, eight sessions, twelve leaders, with ongoing advisory after rollout.

Role: Technical Project Manager · Transportation company (contract) · May 2026–Present

05 / Outcome

The support Rovo agent now resolves offshore support questions natively from live Jira and Confluence context, cutting senior architect escalations by 40% — and it only shipped once it cleared the 95%+ accuracy bar, not on a launch date. The JSM intake analysis (42% needing 3+ exchanges) turned into a build-ready ticket: field additions, helper text, and separating automated alerts from the human queue.

TICKET INTAKE, BEFORE/AFTER — DATA VISUAL

FIG 02 — Comment exchanges per ticket, before and after the JSM redesign.

06 / What I would do differently

I’d have written the PR-gated guardrail before the first agent shipped, not after — the unsafe override existed because governance was retrofitted onto something already running. Now it’s the first thing I build, not the thing I fix.

← All work7 case studies Next / W02 →Intranet at National Scale