Experience
- AI Agent Platforms
- Evaluation Workflows
- Retrieval & Tool Use
- Throttling & Reliability
- Python · TypeScript
- Technical Program Leadership
- Cross-Team Delivery
- Roadmaps & Milestones
Leads delivery for a program of 10+ engineers across squads building an AI-powered risk-analysis agent platform used daily by Azure engineering and SRE leadership. Writes scanner rules for throttling and retry anti-patterns, and builds evaluation workflows so agent output is verified before it is trusted — technical program leadership across squads as much as hands-on engineering.
- 8 static-analysis scanners / scanner rules in production pipelines
- 253 audited detections at a 91.7% true-positive rate
- False-positive rate tuned from 3.2% down to 0.4% through hand review of findings
Key takeaway: An agent's output is a claim, not a result. Verification is what makes trust scale.
Technical lead for enterprise engineering teams on complex production issues across compute, networking, storage, and data, through to resolution. Built knowledge bases and best-practice guides that scaled across a large customer base, and turned recurring pain points into prioritized tooling improvements with product and engineering teams.
Key takeaway: Enterprise problems get solved in rooms before they get solved in code. Align the people, then fix the system.
Diagnosed and resolved global production incidents in distributed customer workloads, and fed recurring patterns back to engineering to prevent recurrence.
Key takeaway: Reproduce it, isolate it, prove the fix. If it keeps happening, it is a pattern, not an incident.
Built full-stack ecommerce experiences in JavaScript for outdoor and specialty retail: discovery, cart, checkout, authentication, and account management.
Key takeaway: Done and live beats perfect and planned. Real customers on real deadlines teach you what shipping means.
Selected Work
A retry-storm simulator plus hand-reviewed scanner findings for retry anti-patterns in public code. The eval-first method, in miniature.
Deterministic, explainable intent-aware re-ranking: same catalog, different intent, different order, across streaming, music, e-commerce, ride matching, and food delivery.
Contract-clause intelligence: semantic clause-vs-playbook matching with calibrated STANDARD / REVIEW / BLOCK verdicts and an automated canary gate for safe rollouts.
Continuous review compliance as a folder in your repo: JSON definitions, a Python scanner, and a CI template. No infrastructure, no SaaS.
More on GitHub: github.com/jtwalters25
Writing
Every claim labeled Confirmed, Inferred or Missing. Precise credit instead of inflated credit, a weekly rhythm, and how to make it safe to share outside your company.
Notes from the middle of the climb: problems over tasks, multiplying others, why impact has a blast radius, and strategy you can explain in a paragraph.
Outcomes without authority, and without becoming the janitor. What a lead actually owns, influence without authority, nudges instead of nagging, and telling the truth upward.
Big ideas are cheap now. Proof still isn't. What a slice actually is, three that taught me something, and when slicing is the wrong call.
Field notes from a failed attempt to make an engineering workflow agentic: a coordinator that spent its whole budget before the A/B test started, why reusing existing skills beat adding agents, and a rule change that passed every test but found nothing in real repos.
Education / The Path
The non-traditional route is the point: bartending before engineering. Code Fellows turned a career change into a craft, and the bar left its mark — read the room, stay calm when it slams, and never serve something you haven't checked. That instinct never left; it is why verification comes first now.
My Philosophy
People first. Always.
Empathy. Synergy. Service. Curiosity.
Most Proud Of
Everything else on this page is downstream of showing up for them and building a life they can be proud of.
Colleagues, managers, and teams hand me the hard problems and the people side of shipping, and trust me to land both. That trust is earned in outages, in reviews, and in rooms where the stakes are real.
A system I designed in early 2026 for translating intent, human or agent, into deterministic, verifiable execution. This year the industry shipped its own versions of the idea. The instinct was right, and the receipts are public.
How I Structure My Time
Illustrative split, not measured data. The order, though, is honest.
Attributes
- Technical program leadership
- Cross-team delivery
- Eval-first thinker
- Calm in an outage
- Ships and iterates
- Raises the bar on verification
- C# depth
- Saying no to side quests