Projects & Tools
A Senior Manager who ships the tool instead of requesting it. Some of these you can open and use right now. The rest run inside T-Mobile, so all I can do is show you how they were built and what changed because of them.
Hartland's assessor releases a "commitment book" every year — every parcel, owner, valuation, and tax bill in town. It's public, it's complete, and it's a PDF. Answering "did my taxes go up?" meant printing two of them and comparing by hand. My family had been doing exactly that.
The real story turned out to be one nobody in town could see: between FY2022 and FY2026 the mill rate fell from $21.80 to $15.85 while assessed values nearly doubled. A resident whose assessment jumped 82% saw their actual bill rise 32% — a revaluation, not a tax hike. Every property lookup now says so in plain language.
Most "take-home pay" calculators apply a flat percentage and call it a day. This one runs the actual 2026 brackets, the Social Security wage base, and the Medicare surtax, then lets you pull each lever — pre-tax 401(k), employer match, ESPP discount, RSU vesting — and watch your real take-home and long-term wealth respond.
The premise is that capture should cost nothing and organization should be someone else's job. You send a message to a personal capture address or type into the app; a Claude Haiku triage function assigns type, priority, life-domain category, and effort size, splits multi-item captures into separate records, and files everything into a hub you never have to sort.
It's running a private pilot right now. If you want in, get in touch — I'm still adding testers.
Most "best places to start a business" content is ranked listicles with no visible method. This runs live Census calls — American Community Survey for population, County Business Patterns for establishment counts by NAICS code — and indexes the result against the national average so the number means something regardless of county size.
This began as a parallel intake route for rush requests that didn't fit the standard process and kept getting stuck in it anyway. Building that made it obvious the standard process was the problem, so I analyzed how the existing system was actually used — 5,492 tickets across 1,671 columns — and found roughly 30 columns carried real signal and 92.7% of volume was a single work type. The redesign was built for that reality instead of the documented one.
Merchandising placement used to live in spreadsheets that were always stale and always a source of escalation during launches. Answering "what's live on the digital surface today?" meant querying several trackers and chasing people in chat. This replaced both.
Most of what's above is internal, so the code isn't something I can show you. The ideas are. These three have each been reused past the project that produced them.
An IP-allowlisting rule between two systems had stalled an automation for months. The standard fix is a network exception request, which is a quarters-long conversation with a security team. Instead I inverted the direction of the call: rather than pushing from the blocked side, I used the ticketing platform's own scripting layer to pull from inside the allowlisted perimeter. The constraint stopped mattering.
When an integration is blocked by a network boundary, check whether either side can initiate — the block is usually directional, not mutual.
Work tiers had become a negotiation. Requesters knew what tier meant for turnaround and staffing, so they optimized their submissions for the tier they wanted rather than describing the work honestly. That corrupted the intake data at the source.
The fix was to stop asking. Tier is now derived from project type, level of effort, page URL, and attachment count, and it's visible only to the internal lead. The requester sees a T-shirt size.
Removing the lever removed the incentive to pull it, and the intake data got honest.
A weekly performance deck was generated from a slot-driven, deck-first architecture: each slide knew which cell it read from. When a section dropped out of the source extract, the slide didn't error — it silently reused last week's numbers. Nobody caught it, because a stale number and a fresh number look identical on a slide.
I rearchitected it onto a canonical long-format model — one row per section, subsection, entity, metric — behind a query layer that raises rather than returns empty, with the presentation formulas isolated in their own layer. The companion validation pass caught four real mapping bugs that row-count checks would have missed, including a semantic subclass bug where "Prime" was matching "Sub-Prime."
Now if a section drops out of the extract, it drops out of the deck. A visible gap is a bug report; a stale number is a wrong decision.