Senior Technical Project & Delivery Manager · PMP
18 years delivering client software from the vendor side: turnarounds, release governance, fixed-price and T&M engagements up to ~$800K. I also build production web apps with AI coding agents, under the same gates I install for clients. This page shows how.
Approach
Two personal products, built since July 2026 with Claude Code as the engineering team. I own scope, architecture decisions, acceptance and release. Every change follows the same path, and nothing reaches production on the strength of its own author's self-check.
The delivery path for every change
Plan item, acceptance criteria, open decisions logged with their rationale.
One AI session builds on its own branch and opens a pull request.
Automated. Red blocks the merge.
A second session that did not write the change reads the diff, reruns the suite and reproduces the claims.
Squash-merge, auto-deploy, check on the real device.
Every design decision with the reason and the alternative it beat. Reversals stay in the log as reversals, so nobody re-argues a settled point.
Append-only record of what was checked, how, and what is still unmeasured. "Verified" has to point at a run, a query or a recomputed number.
A quality rule without an automated check is a request. Each check ships with a counter-test proving it can actually fail.
Case study 1
Problem
A personal collection of 250 recipes lived in Notion. Notion is slow to navigate in the kitchen, does not work offline, and has no cooking view. The goal: one user, two devices, portrait and landscape, every core task in three taps or fewer.
Approach
Governance
Outcome
Migration: flag it, don't drop it
The Notion export runs through a rule-based converter that lists anything it cannot parse instead of silently skipping it. The first run reported 250 recipes, 250 photos and 12 items to review. Eleven of those had the same cause: step photos indented under a step. That was a structure the converter did not know, and a silent converter would have lost them without anyone noticing. After one fix the second run brought in 335 photos with nothing left to review.
Screens from the design prototype with sample data. The app's interface is in German for its user.
Case study 2
Problem
Calorie counters treat 100 calories of lentils and 100 calories of candy alike. NutriTally weights each item's cost by nutritional quality, using the published Nutri-Score 2023 algorithm, inside a daily and weekly allowance. The domain is unforgiving: one misread rule or one blank treated as zero changes every number the user sees.
Approach
Governance
Outcome
When the review gate lapsed for a day
Under time pressure I let one session build, self-check and merge a day's work without independent review. It shipped a swipe-to-close gesture, carefully guarded against scroll conflicts. Nobody asked what that gesture throws away. Shortly after, a meal's worth of selected ingredients vanished with a swipe. A cold reviewer asks "what state does this discard?" first. The author does not. Since then the review gate binds every change, small polish items included, and if no reviewer is available the work stops at the open pull request.
Five fixes for the wrong problem
"The screen floats" took five rounds of fixes to page height. Each fix was real and each measurement correct, but none addressed the report. The cause was iOS Safari zooming into text fields smaller than 16 px and staying zoomed, which desktop browsers never show. It surfaced only when I, as the user, described the symptom again from scratch instead of answering "does it still happen?". Now a second round on the same symptom means re-checking the assumption first, and the 16 px rule is an automated test in both apps.
Screens from the Ume redesign prototype with sample data. NutriTally is a personal learning project and does not provide medical, nutrition or health advice.
What carries over to client delivery
A review counts when the reviewer did not build the change and works from the branch, not from the author's summary. The same holds for human teams.
One day, a CI wait loop accepted an empty list of checks and two changes merged with no CI at all. Every gate now names the condition under which it goes red.
Agents produce correct code for the problem they are given. Framing the problem, asking what a change throws away, and deciding when to stop are still the delivery manager's job.
Decision and QA logs kept during the work, not after it, let any new session or team member pick up the thread without re-asking settled questions.