AI products
Figure out what a model is actually good for, build the thing people use, and decide when it’s ready.
05 About · updated October 2026
Ben Schippers I build things, mostly with Claude, and write down how.
Cloud engineer at Microsoft by day. Tree-ring scientist before that. Since November 2025, products and small science experiments at night, built with AI. This is the longer version.
Advanced Cloud Engineer at Microsoft · Atlanta, Georgia
What I believe
Abundant execution, scarce judgment
Frontier models changed what one person can build. Choosing the problem, designing the system around it, catching the lie, keeping the miss, and deciding what earns release are the same job they always were.
More work runs in parallel. None of the judgment moves.
- Frame
- Find the problem the rest depends on, and define what good means before anything is built.
- Orchestration
- Give people and agents the context, boundaries, interfaces, and quality gates to move in parallel.
- Release
- Show the work, narrow the claim to what was measured, and stop the release when a check fails.
What I like doing
Three kinds of work
The work I keep coming back to, at Microsoft and at home.
Big technical programs
Keep a lot of parallel work moving without losing the thread: who owns what, what blocks what, and what could go wrong.
New, messy problems
Walk into something nobody has figured out yet, listen to the people living in it, and carry it from first sketch to something that runs.
Things I’ve built
Four stories, with the numbers
A team at Microsoft, a tool anyone can install, a live product, and the way I run the lab at home.
Case 01 · at Microsoft
Microsoft · from silos to signal
Turn fragmented customer signal into a shared system that changes engineering priorities.
- Problem
- A major enterprise AI rollout was accelerating across product lines with no shared view of the adoption walls customers were hitting.
- What I did
- Assemble the cross-functional forum, deploy real-time case analytics, and turn fragmented support signal into a durable engineering-priority loop.
- Direct result
- The system surfaced 76 blockers and directly unblocked 7,380 users.
- Program outcome
- The broader work helped the organization add approximately 94,000 seats. That is a shared program result; my lever was the pipeline and the forum that ran on it.
Case 02 · open source
Career Compass
Publish the tool where a stranger can install it, then aim the audit at what they would actually receive.
- Problem
- A local-first tool that holds a person's contact details and salary expectations also has to read job postings written by strangers. Both end up in the same context window.
- What I did
- Ship it publicly as a TypeScript MCP server on Anthropic's Model Context Protocol, installable in one command, and point the audit at the published package rather than the repository — driving it over the wire the way an installing stranger would.
- Quality gate
- In a July 2026 audit pass, 57 agents ran five independent lenses over it, and every non-speculative finding went to a verifier prompted to refute by default. 52 findings were raised, 36 survived that verifier, and 16 were killed.
- What the audit changed
- Silent data loss on the happy path, measured against the published server: 8 of 8 runs dropped records before the fix, 8 of 8 kept them after. Untrusted posting text now sits behind a nonce fence at all 10 interpolation sites. Each guard test was checked to go red when the thing it guards is broken.
- In use
- Its one external user to date, a laid-off job seeker, ran it through an active interview process. She reported that its value was a fit verdict that contradicted the job board's own labels in both directions, and she used it to project an interviewer's remaining questions mid-process — a use it was not designed for, which redirected the roadmap. Shared with her consent.
Case 03 · a live product
Visual Inventory · every change through one gate
Run a live, billing-enabled product where nothing reaches users except through the same gate.
- Problem
- A product that catalogues what is inside people's homes has to keep one household's data out of another household's reach while still shipping changes.
- What I did
- Launch with billing enabled, then route every change after it the same way: a numbered pull request with CI gating, including my own.
- Quality gate
- Since the 2026-06-19 launch, all 64 merged commits arrived as numbered pull requests with CI gating, and none as a direct push. Across the repository's whole life, 622 pull requests have merged, against a suite of 2,268 tests, 2,266 of them passing on 2026-08-08.
- What an audit found
- A post-launch pass found a cross-household exposure in the database policy layer: the profiles UPDATE policy carried no WITH CHECK, so an authenticated user could reassign their own household and read another household's inventory. It was closed with a guard trigger proven across all five caller contexts. The product's readiness contract, clauses S1 through S6, holds that it does not call itself ready until a fresh production pass flips them.
Case 04 · how the lab runs
The lab's operating system
Put the operating rules into programs that refuse, and let them report where the record and the system disagree.
- Problem
- A fleet of agents produces more work than one person can review by reading all of it.
- What I did
- Sort every action an agent might take into three tiers: work it may merge unsupervised, work it must surface and wait on, and work it never takes in any mode. Nine conditions have to hold at once for the first tier, and eleven actions are always held. The default under ambiguity is to hold, and every unsupervised merge is logged in the response. The rule is written as “no surprise merges.”
- Quality gate
- The agent doing the work is never the one grading it. Two of the four verification rungs are programs that refuse: the ledger tool will not mark an item done without an attached commit, pull request, or URL, and a session-end gate blocks the wrap while today's session record fails its linter. That linter is wired a second time into the automatic commit, which unstages a failing record and leaves it visibly in-flight. Both of those gates fail open on their own breakage. The other two rungs are written contract.
- The record
- 442 session records over 101 days as of 2026-08-08, each required to name what shipped and what is still open. An inventory of 228 entries carries intent and provenance on every one, and a scanner re-derives the live facts and may only report disagreement, never edit the record: 6 flags on 2026-08-08, about 2.6%. A drift report regenerates at 07:00 daily, from a scheduled task running since 2026-04-26, acts on nothing it finds, and refuses to emit a clean report it could not substantiate; silences are named files that can carry an expiry. The machinery currently reports warnings and drift against itself. The claim is that the disagreements between the record and the system surface on a schedule, whether or not anyone wants to see them. It is not a claim that the lab is clean. These figures are measured by the lab's own instruments and are not externally verifiable.
The lab
Ground rules
Agents get enough rope to search, branch, build, test, and challenge in parallel. Accountability stays legible from the first framing decision to the final merge.
- 01Frame
Find the load-bearing problem and define the change that would matter.
Human-owned - 02Bound
Set context, interfaces, the checks a result has to pass, and refusal conditions.
Human-authored - 03Run
Let agents search, implement, test, and report through parallel branches.
Agent-executed - 04Challenge
Use adversarial review, counterexamples, and independent checks against the claim.
Shared scrutiny - 05Release
Merge, publish, narrow the claim, or refuse the result. Preserve the miss.
Human decision
The one rule
The fleet can propose, build, test, and run through the night. It cannot decide what deserves to be true in public. That decision stays mine.More on how the lab runs
Career
Eleven years around Microsoft, five inside it
The Microsoft years are the record at organizational scale. The lab is the same work smaller and faster, left in the open.
Enterprise operating record
5 yearsas a Microsoft Senior Program Manager across Copilot, Graph, Windows 365, and Teams Devices
8product lines in an internal platform portfolio spanning signal, routing, self-service, and quality measurement
95+features shipped through a rebuilt signal-to-engineering pipeline
433Kusers in a crisis transition led with near-zero churn
3 → 150agents in the Premier Engineering program built from pilot to scaled operation
60K/yrincidents handled by that scaled program
Self-reported career figures, not publicly verified. Detail available in conversation.
Read the full career cross-sectionFleet pulse
Git-derived, timestamped, and linked to the public repositories. Hidden when older than three days, the same tolerance the home page uses.
—private PRs merged · 30 days
—private repos with merged PRs · 30 days
Private fleet, anonymized — same query family with is:private. Nothing outside the lab can confirm these, by design; repo names never appear.
Background
Raccoons, tree rings, and a $1,000 credit
Start with the website. A fish in the tank surfaces only between midnight and 3 a.m.; it knows when you’re up too late. A popsicle-stick cottage stands or falls under working physics, depending on how you glue it. On the windowsill, green means human-promoted, amber means a machine check awaits review, and a folded grey leaf keeps the miss on the plant. A computer in my house waters it with science while I sleep. Anyone can leave a pebble in the soil.
None of this was on a roadmap. November 2025: I had never made a single GitHub contribution. I picked up a $1,000 credit for Claude Code with one goal: see how fast a couple of side projects could burn it. The first revelation was velocity. The important one was responsibility. When execution becomes abundant, judgment becomes the bottleneck. I hold the problem, the context, the quality bar, and the decision about what earns release. We have a lot of fun.
The kid with the microscope and chemistry set grew into a dendrochronologist handling field samples in a tree-ring lab, and never really left. The same instinct later ran at organizational scale: Microsoft, enterprise AI, multi-team platforms, signal systems, and operating mechanisms that moved customer friction into engineering action. Broken Branch continues that work under frontier conditions.
The raccoons ride along in the long way, a road-trip companion with three of them narrating the drive. I still make the first-time face every time the lean-to cottage takes its own weight. Cute gets a stranger through the door; then the physics has to hold. Try the labs.
The plainer arc: eleven years in the Microsoft ecosystem — two at Mural migrating partners onto Office 365, four building Microsoft Premier through Experis, five directly at Microsoft — a 10-month gap where shipping in public became the job, and back at Microsoft since March 2026.
Career
Microsoft
– presentAdvanced Cloud Engineer
Returned March 2026 — now on the cloud-engineering side of Microsoft's enterprise AI work. Program manager by training; engineer by current title.
Microsoft
–Senior Program Manager, AI Platforms & Enterprise Operations
Across 5 years — Copilot, Graph, Windows 365, and Teams Devices among the 8 product lines
Owned a portfolio of internal platforms across 8 product lines—signal systems, routing intelligence, self-service tools, and quality measurement. The infrastructure that turned customer friction into engineering action.
- Rebuilt the signal-to-engineering pipeline from scratch. 95+ features shipped through this system; adopted org-wide.
- Built routing intelligence that classifies incoming work by complexity and matches it to the right skill level.
- Scaled self-service from pilot to ~50% adoption on flagship products. Tens of thousands of tickets per year that never get created.
- Created early risk detection identifying 700+ at-risk customer situations before they escalate.
- Built and shipped a recommender reaching 14K enterprise customers with 14% conversion.
- Led crisis response for a 433K-user transition—near-zero churn.
Experis / ManpowerGroup
–Program Manager → Operations Manager
Microsoft Premier Support
Built Premier Engineering from a 3-person pilot to 150 agents handling 60K incidents/year.
Mural Consulting
–Engagement Manager
Managed partner programs spanning 1,000+ Office 365 migrations across 12 global partners.
Education
B.S. Interdisciplinary Science & Technology — University of Arizona
Former dendrochronologist. Yes, tree rings. It's where the domain name comes from.
Case study: Copilot extensibility, from silos to signal
Led the cross-functional effort to build enterprise AI adoption intelligence across multiple product lines. The signal-to-engineering pipeline I built surfaced 76 blockers and unblocked 7,380 users directly — work that helped the org add ~94,000 seats. (Program-level outcomes; my lever was the pipeline and the cross-team forum that ran on it.)
76 blockers found in this case
7,380 users unblocked in this case
94K seats added
47% self-help success
A major enterprise AI rollout was accelerating across multiple product lines with no shared visibility into adoption patterns. Customers were hitting adoption walls that no single team could see. I assembled a cross-functional team, deployed real-time case analytics, and built the feedback loop that turned support signal into engineering priorities.
By the numbers, Microsoft era
3 → 150agentsCo-founded the program, hired the team, built the playbook
700+blockers · career totalSurfaced from support signal that product teams couldn't see
95+features shipped64% of requests submitted to engineering accepted
220Kusers unblocked · career totalAdoption walls removed before they became churn events
94Kseats addedProgram-level outcome; customers who expanded after we resolved their blockers
8product linesOne internal platform portfolio spanning signal, routing, self-service, and quality measurement
Read “Growth Rings” — the 10-month data story →
Past work. Scouts+: adult merit badges with real requirements, local troops, verification, and physical rewards. Full stack with Stripe and nineteen database tables. Live, signups closed, winding down.
Contact
Microsoft has my weekdays. Write if you’re building something with AI and want a second pair of eyes, or if you just want to compare notes on what these models can do. Client details stay off this site.