Accounts of specific engagements: what the situation was, what was built, and what resulted. Clients are not named. Figures are either exact or labelled as estimates.
Part one
Land and entitlement.
Site development is documentation-heavy work that ends in a public hearing. It rewards knowing what is true before you have to defend it in front of a room.
Speed under deadline · Where the practice started
317 packages, 60 days, three days to build the pipeline.
A national wireless carrier issued a relocation portfolio and gave 60 days to return feasibility and permitting recommendations across it. Assembling that by hand was not achievable in the time available.
The carrier’s site list was synthesised with spatial data: coordinates, ground elevation, and urban or rural classification driving how large each search area needed to be. Python generated a review radius per candidate on that logic, half a mile urban and one mile rural, and a macro-enabled Word template produced the folder structure and a pre-populated feasibility analysis for every site.
Each generated document arrived already carrying its site number and name, state and jurisdiction, coordinates, ground elevation, existing tower height and structure type, carrier centerline and height above ground, and review radius, plus structured sections for search-area feasibility, permitting paths split between raw land and colocation, feasibility maps, and the relevant zoning ordinance excerpts.
317 feasibility packages across 315 sites, delivered on schedule. The pipeline was built in three days by an operator with no prior programming experience, and the analyses were divided across the three people working the programme.
The analysis used four attributes. Coordinates, structure height, city, state. That was what the spatial logic required and it is all the pipeline ever touched, decided before the code was written rather than reviewed afterwards.
The three-day build is corroborated by file timestamps, not recollection. Artifact names and client identifiers withheld; detail available on request.
Date
Artifact
What it establishes
Oct 7, 2024
Executed scope of work
Engagement scoped
Oct 22, 2024
Carrier site list received
Source of truth established
Oct 22, 2024
KML containing 315 review-radius polygons
Python radius generation working
Oct 23, 2024
Folder-generation macro
Directory structure automated
Oct 24, 2024
KML containing 315 site placemarks
Candidates plotted
Oct 25, 2024
Production macro-enabled template
Pipeline complete
Stated as an estimate: the labor hours this removed run to several hundred. That figure is a considered estimate, not a measurement, and it is not presented as one. The delivery counts are exact: 317 packages, 315 sites, a 60-day window, a three-day build.
Source of truth. The carrier’s own site list, treated as authoritative rather than re-keyed.
Structured operating data. Every candidate as a geospatial point with a market-classified review radius.
Workflows and controls. Radius generation and document assembly automated, with one template governing every output.
Tested and handed off. 317 packages delivered against a fixed deadline.
Influence
Documents that became the standard.
Two national carriers independently took documentation built for internal workflow and adopted it as their own enterprise standard.
One deployed the feasibility and site documentation forms across its national vendor base. The other adopted a site candidate information package as its internal standard. Neither was pitched a product. Both encountered a working document and made it the bar their other vendors had to meet.
The follow-on work proved the forms were doing real work rather than formatting it. Quality review of other vendors’ submittals against those same forms reached different feasibility conclusions, and recovered sites those reviews had ruled out.
Both carriers are named on request.
Mapped to the method
Source of truth. One controlled form defining what a complete site record contains.
Structured operating data. The same fields, in the same order, on every submittal from every vendor.
Workflows and controls. A review standard that could be applied against work the author did not produce.
Tested and handed off. Adopted enterprise-wide by two organizations, and operated by their vendors rather than by its author.
Standing in the room
Several hundred hearings. Three denials in twelve years.
Site development ends in public. Zoning boards, plan commissions, boards of adjustment. I have presented several hundred times and been denied three.
One of those three I called before we filed. The client wanted to proceed at a site that did not meet the ordinance and I told him we would lose. The other two were decided on political rather than regulatory grounds, at sites that met every published requirement.
That record comes from doing the jurisdictional work before anyone commits capital, and from not bringing a body a case it cannot approve.
Stated precisely: the denial count is exact, three. “Several hundred hearings” is a considered estimate rather than a counted figure.
Source of truth first. Title, easements, ordinance, staff position, and findings resolved before the application is filed. The homework is the whole job.
Contested work
Two contested outcomes.
Two situations where the straightforward path had already been ruled out.
The blanket easement. A rural build parcel everyone had written off. A Canadian wind energy company had told the landowner its easement would cover fifteen feet on either side of its trunk line, then recorded a blanket easement across the entire eighty-acre parcel, rendering the property unleasable. Rather than move the search ring, the issue was worked across an international border until the developer agreed to re-survey and amend the agreement down to its actual footprint. Title cleared, ground lease secured, tower permitted and built.
The contested hearing. A national tower company learned a competing permit application was being filed and sent outside counsel to argue against it. Every hearing material, including decks, exhibits and supporting findings, was built in house. The application was approved over the objection of opposing counsel.
The relocation automation above belongs to the land practice: it was written inside infrastructure work, to deliver an infrastructure portfolio. It is where the software started, not where it stayed. Everything below was built after October 2024, deliberately in sectors outside telecommunications, and is running now.
Seven operational systems, in daily production use
Built for a multi-property operator: expense reporting with bank-statement import, receipt OCR and a quality gate that blocks submission on failure, in production since the end of May 2026; a property and location directory that every other system resolves against; linen inventory across five property trackers with a controlled master catalog; a mobile consumables inspection application with photographic evidence and duplicate-proof submission; a production inspection tracker in daily use; a turnover and issue tracker; and an authenticated employee portal that ties them together.
Engineering controls: versioned source packages, schema versioning, staged QA gates, deployment preflight, rollback planning, stable identifiers, protected formula columns, approval logs with approver and timestamp.
Operational result: a five-person team runs its month on these systems, without a technical specialist among them. The portal is live behind Workspace sign-in, and its protected interfaces fail closed to anonymous requests. Stated as an estimate: building one tracker by hand ran to more than two hours, and onboarding a new client to more than forty.
Delivered as applications, not documents. The consumables inspection runs as a progressive web application opened from a durable QR code at the property: offline once loaded, drafts staged on the device, a photograph required against each of 28 governed items, sign-in demanded only at submission, and repeated sends made idempotent by client submission identifiers and a server-side receipt registry. A client intake and onboarding application is now a deployed private pilot, owner-only, with live email, contract, signature and property-record writes deliberately disabled until their approval gates are complete. One intake path, from first conversation to property setup →
A field inspection committed successfully, then a slow path returned an error to the operator. The wrong move is telling a user to retry. That is how you get a duplicate record in a system people trust.
What happened instead: the transaction was identified by its original client submission ID, submission ID and payload hash, confirmed as already committed, and recovered without duplication.
Recovery paths of this kind are designed in advance, because the alternative is discovering the gap during an incident.
Scope: project control, on the Mac and the phone
A native macOS and iOS application for time tracking, project control and invoicing, sharing code and a data model across the two platforms. It runs a single clock and requires a sitting split across projects to sum back to the sitting, so the same hour cannot be billed twice.
State: Apple accepted the first notarised build 4 hours and 24 minutes after the first commit, and 18 accepted notarisations in all across the twenty-five hours that followed, averaging one every 88 minutes: the original dated record. 87 commits and 497 tests at that point; 869 Mac tests were passing on 8 October 2026 (879 in total, 10 skipped). Version 1.0 (6) is in internal TestFlight beta and not publicly released. There is also a verification script that plants a defect and requires its own detector to catch it before a pass is believed.
A native macOS application for working across a knowledge base, projects and workspaces, meeting processing, review, backup and restore, with provider-flexible AI under human approval.
State: installed and in active internal use, in continued development. Signed and notarised by Synergistic Solutions Unlimited, and accepted by Gatekeeper as a notarised Developer ID source. Its test suite is roughly the size of the application itself.
Not commercially available. It is installed and used internally, and not distributed.
Lucid: native macOS writing application
A writing assistant that works inside the other applications you already write in, with every suggestion shown as a tracked change you accept or reject, and your keys kept in the system Keychain. It works with the AI you choose, including private options that keep your writing on your own machine. Only the options observed working end to end are claimed.
State, measured on 28 August 2026 at version 0.1.17: installed, Gatekeeper-accepted as a notarised Developer ID source with a valid ticket. 19 submissions accepted by Apple’s notary service. 312 tests executed with 0 failures, measured on 28 August 2026. Tests that could not run say so by name rather than passing quietly. Not publicly distributed, and nobody other than me has run it.
What using it actually changed was the review process, not the software. On both occasions I have used it on real work, I found something no review had, including a panel that could not be dismissed by any documented path. A Gate-2 review had flagged that and ranked it cosmetic. Contact with a user was what corrected the severity judgement.
Shared Universe: one codebase, five surfaces
A private family media catalog on iPhone, iPad, Apple TV, Mac and the web, from one shared codebase and one backend.
State: Build 63, in TestFlight for iOS, tvOS and macOS, in daily use by one household. Private, and not for sale. Five surfaces, one catalog →
RepDial: strength training, on iPhone and Apple Watch
A Tempo Dial coaches every rep, with voice and Apple Watch cues, plus programs, progress and nutrition.
State: public TestFlight beta. 765 commits from 25 September to 8 October 2026, 1,084 package tests plus 776 app tests passing, and 54 builds uploaded. Apps in development →
GatherRelay: family care, kept in one private place
One calm place where a family keeps care paperwork and shares only what each person needs.
State: private beta, live on Cloudflare since 7 October 2026, and coming soon to the App Store for iPhone and iPad. gatherrelay.app →2,971 automated tests, and 41 database test suites with 2,707 assertions, none failing, as read on 8 October 2026.
Glimmer Hunters: a learning adventure for iPad
Reading, feelings, logic and music, with 40+ mini-games across four lands, for children aged about 4 to 7.
State: in testing on iPad. Captions on by default, a visual and haptic version of every sound game, and no ads, no tracking and no subscriptions. Optional expansions are only ever bought by a grown-up. glimmerhunters.app →
Websites, built and checked from one source
Nine web properties are in production, including this site and the three apps for a property management client. Each site is generated from one source, and a checker blocks drifted figures, dead links and uncleared names before a page goes out.
State: that client’s public site has been rebuilt as 23 pages with 277 automated tests, and its checker catches 18 of 18 planted defects. It is in private preview, and launch is not yet scheduled. Website development →
Recent work
Five accounts from October 2026.
01 · A property management client
Three live apps moved to the client’s own cloud in one evening
An employee portal, consumables and inspections, moved from a hosted builder to the client’s own Cloudflare account.
Data copied and counted. Every secret regenerated straight into the keychain and never displayed, and a publicly served build manifest closed along the way.
02 · GatherRelay
Bugs that only existed in production
Code that passed every Node test failed on Cloudflare’s runtime. A test harness now runs on that runtime and catches that whole class of bug.
Separately, a hosted-only test caught an over-broad database permission on day one, and a migration closed it.
03 · Glimmer Hunters
Accessible from the first build
Captions are on by default.
Every sound-based game has visual and haptic equivalents, so children who are deaf or hard of hearing can play.
04 · How this practice runs
Running a crew of AI builders on one machine and a fixed budget
A dozen AI builders worked at once across several products. They shared one computer, two usage budgets that reset on different schedules, and one person who approves anything public or costly, often while he is asleep.
What was built: a way for many AI builders to share one machine and a fixed budget safely, keep working while they wait for an approval, and learn from each other’s mistakes the same day.
Result: several products shipped test builds in parallel, and one app went from internal build 6 to build 17 in about 19 hours.
69% of one coordinator’s incoming messages were acknowledgements and completion claims. Reporting overhead is a cost you can measure and cut.
A test run reported success because its last step was a print statement. Read the result, not the wrapper.
05 · How this practice runs
A tool should report what it recorded, not guess from what is missing
In one night, the system that coordinates more than a dozen automated workers produced three failures of the same kind. Each time the work was done and the report was wrong.
One report said delivered mail was still waiting, and a lead spent the gap hunting mail that was never lost.
Questions waiting on me were marked answered while I was away, in words I never wrote.
One check confirmed that a message had been read when it had no way to know.
What changed: reports now state only what was recorded, and a wrong answer can be corrected without being lost. For a client, this is the dashboard that reads green because the data never arrived.
A privacy-first communication and care-coordination system, built under pressure and since rebuilt as a fully synthetic public case study. It shows ambiguity being turned into clear roles, reliable workflows, risk controls, and a humane experience for people who had no capacity to learn a tool.
Product strategy
Program leadership
Systems thinking
Privacy by design
Rapid execution
Stakeholder communication
The origin was not a brief. My own family needed one private place to keep the people who cared about us informed, and the communication overhead had become its own operational problem: one person answering the same questions repeatedly, correcting information that had already gone stale, and coordinating offers of help across a large network while circumstances changed by the hour.
Social platforms were the obvious answer and the wrong one. They put delicate information where it can be misconstrued or reach people it was never meant for, and they set the terms of audience and permanence on your behalf. The alternative was to adapt the need to the tool.
The sequence mattered more than the feature list. Private access, dependable official updates, clear publishing confirmation and duplicate prevention came first, because reliability precedes novelty when people are relying on a single source of truth. Vague offers of help then became claimable meals, visits, rides and errands. Capability-scoped roles let trusted organizers coordinate without receiving contact details or administrative access. Navigation, unread badges, refresh behavior and notification preferences were changed later in response to observed friction, not anticipated preference.
A later revision removed visible photo descriptions from the interface while retaining meaningful alternative text for screen readers, because some of the people using it depend on that.
Risk was controlled structurally rather than editorially. The public case study and the private system are separate builds with no shared data plane: no production database, no real authentication, no messaging, no private media. The demonstration shows the method without asking a redaction filter to protect a live family system, which is the same judgment that produced the private application in the first place.
It reached a working private application in two calendar days on roughly six hours of hands-on time, and it is in daily use by the people it was built for, my family among them. It has kept changing as the need has changed.
The need has always directed the development. Not the reverse.
Presented publicly with no real identities, messages, photographs, or production connections.
A technology-transfer program is designed and not yet delivered. The three Mac applications are signed, notarised and installed, and none of them is distributed. They are in internal use, they are not for sale, and there is no download. None of the six native applications is on the App Store yet.
Every figure above is either exact or labelled as an estimate. Where no number exists for hours saved, user counts, or precise production dates, none is given.
