Staff Technical Program Manager · AI-native engineering · San Francisco Bay Area
About 15 years leading large technical programs across developer productivity, build and release infrastructure, and incident management at Snowflake and Flow. Today I build open-source Claude Code skills that automate the mechanical parts of program management, so the judgment calls stay with people.
7d → 1d
Release cycle cut from weekly to daily (Snowflake Ship Faster)
10×
Developer inner-loop productivity gain (Snowflake Type Faster)
7 + 1
Open-source AI TPM skills, plus a harness that tests them
~15 yrs
Leading programs across dev productivity, infra and release
Centerpiece
Four connected skills take one PRD all the way to an executive dashboard. Each one passes structured JSON to the next. In every skill, Claude makes the judgment calls and plain, tested Python does the math, so every number is reproducible.
What the pipeline caught in the worked example
One synthetic PRD ("Nimbus Expense") becomes 5 epics, 23 tasks and 78 story points, with a planned finish of 2026-08-20. The status report rates the program RED because the critical path slipped one day, to 08-21. The Monte Carlo forecast reaches the same 08-21 P50 by a completely different method. It also finds that the team that caused the slip is now on track, and the real tail risk sits with Backend. That second finding is the kind of insight nobody gets when four reports live in four documents.
Selected work
All projects are open source under the MIT license. Every example uses synthetic data and no employer information.
01 · Plan
Turns a PRD into a Jira-ready backlog with epics, tasks, cross-team dependencies and Fibonacci estimates. It replaces a couple of hours of manual breakdown per PRD.
5 epics · 23 tasks · 78 pts 4 tasks flagged high-risk
02 · Schedule
Turns a backlog into a real schedule using the critical-path method: early and late starts over business days, slack per task, a Mermaid Gantt chart, and a risk register of every dependency that crosses a team boundary.
16 cross-team handoffs 2 of them on the critical path
03 · Report
Combines the schedule with raw team notes to produce an executive status report. The RAG rating follows an explicit rubric, so real slips aren't softened to amber and harmless delays don't turn the project red.
RED critical path slipped 1 business day → 08-21
04 · Forecast
Checks the plan's assumed team velocity against what each team has actually delivered. It runs 10,000 Monte Carlo trials per team, resampling real sprint throughput, and returns P50, P85 and P95 finish dates.
LIKELY LATE Backend P50 08-21 · P95 08-24
05 · Synthesize
Combines all four outputs into one dashboard. It never recalculates upstream numbers. It validates and merges them, and Claude's only job is the executive summary that connects them, because the most useful findings come from comparing reports.
One page backlog, schedule, RAG and forecast together
Reliability
A test harness that checks all seven skills against their own guarantees, such as slack formulas, percentile order and date order. It then deliberately corrupts known-good output to prove each check actually fails when it should. A check that never fails is reported as a harness gap.
47 checks CI-ready exit codes
Developer productivity
Reads a codebase and writes the briefing a senior engineer would give a new hire: an architecture map, a ranked list of files to read first, gotchas with evidence, and scoped first tasks. Tested on three real open-source repos.
25 stale links found in a real cookbook README
Incident management
Turns a Slack thread, logs and alerts into a postmortem with a timeline, root cause, severity and owned action items. Every duration is recalculated from raw timestamps instead of trusted from the model.
43 min outage not "about 40", the number people tend to remember
AI cost observability
Answers "where did my Claude Code tokens actually go?" It spreads each request's cost across everything in context, so per-item costs add up exactly to the session total. Runs locally with no network calls.
37% of my spend was tool results
Automation
Bulk-archives ChatGPT history by date or keyword from your own logged-in tab. Dry run by default, rate-limited, and re-checks every archive because the API can return 200 OK without saving the change.
"200 OK" ≠ done the lesson that led to skill-verifier
How I work
I use Claude Code daily for planning, analysis, writing and building. I treat it like any other production system: with guardrails, verification and a cost ledger. Here are four weeks of my own usage, measured with token-ledger.
Models handle the tedious parts: decomposition, extraction and first drafts. Decisions that carry accountability, such as priority, risk and go/no-go, stay with people, and the output says which is which.
Every skill splits the work. Claude classifies and explains, and deterministic, testable code calculates dates, slack, percentiles and durations. A number I can't reproduce doesn't belong in an exec report.
Outputs are JSON with schemas, validated before rendering, with evidence attached to each claim. Automation should reduce toil without hiding important decisions.
Experience
Staff Technical Program Manager · Sep 2025 – May 2026
Senior Technical Program Manager · Sep 2022 – Mar 2025
Technical Program Manager, Product Release · Mar 2021 – Aug 2022
Founding Technical Program Manager · Sep 2018 – Feb 2021
Senior TPM, Project Services · 2014 – 2018 · Lead Deployment Engineer · 2010 – 2014
Education & certifications
Toolkit
I'm looking for Staff TPM roles where AI changes how engineering gets delivered, and where better execution systems are the leverage.