FINANCE // TECHNOLOGY // DESIGN
I closed the books. I did not sponsor the automation — I wrote it.
Ran journal entry accounting for a national health system — 68 hospital campuses, 300+ sites of care, 49 accountants. The deterministic Python behind journal entry intake is mine — rolled out, not yet nationwide. As of September 2026 I build close software full time.
No vendor. No statement of work. No IT ticket rotting in a queue. I have shipped production code since 2008 and saw no reason to ask permission to do my own job. I spent twelve years watching the same five days come apart in the same five ways, first at one hospital and then across three hundred sites of care. I am building the thing I kept wishing somebody had already built. The rest of this file is the evidence that I can.
ACT IWho is talking
Sections 01 and 02 — the job nobody has a clean title for, and the method underneath it.
01 PROFILE
It is the one where the accounting is wrong because the system is wrong, and fixing it needs someone who understands the ledger, the operation that produced the transaction, and the code underneath — at the same time, in the same head.
Normally that takes three people and a translation layer. A controller who knows what the number should be. A business analyst to write it down. An engineer to build something approximately like what was written down. Six months later everyone agrees the requirements were misunderstood.
I am all three, which is the only reason I know how badly those roles usually understand each other. I have closed the books and written the software that closes them. I have run the P&L and written the code on the same problem in the same week.
Give me a broken process and I will tell you whether it is an accounting problem, an operations problem, or a software problem — and then fix it, whichever one it turns out to be.
THE USUAL ANSWER
WHAT ACTUALLY HAPPENED
Sponsored an automation initiative.
Identified the problem, designed the workflow, wrote the automation, validated the accounting logic, and deployed it.
Gathered requirements for the developers.
Already had the requirements. I was the one doing the work.
Built a financial system from a technical specification.
Built it knowing what the trial balance was supposed to look like when it was done.
Escalated to IT and waited.
Shipped it Tuesday.
What I do not do is write a deck about what should be built and leave. I have never delivered a recommendation I was not willing to implement myself, and I am not starting now.
02 OPERATING SYSTEM
The jobs look unrelated on paper — hospital finance, state government software, industrial operations, national signage, a trail app. They are the same job. This is the loop, and it has not changed since 1998.
Usually one everybody stopped seeing years ago, because working around it quietly became the job.
Not broken enough to rip out. Just broken enough to eat four hours of somebody's Tuesday, every Tuesday, forever.
Because I have done the work by hand. Requirements defined from ownership, not from an interview.
What it does to the close, the control environment, the audit, and the P&L. In that order.
Well enough to build the better version rather than describe it to somebody who will build something adjacent.
Working software, on real data, usually before the meeting to discuss it has been scheduled.
This is the step everyone skips. It is also the only one that determines whether any of the others mattered.
Usually 01 and 07, or 05 and 06. The value was never in any single step — it is in running all seven with no translation layer in the middle, because that layer is exactly where the requirement quietly turns into something nobody asked for.
Accounting is not really about arithmetic. It is about verification — refusing to accept a number because it looks right, tracing it to a source, and knowing precisely how confident you are entitled to be. That is the same discipline that separates software which demos well from software that works, and it is the thread running through every unrelated-looking job in this file.
It shows up in odd places. Building the trail app on this domain, I checked a state agency's published coordinates against its own published mileposts and found entries that were physically impossible — points whose straight-line distance exceeded the trail distance between them. Two more had transcription errors that would have put a rider in the wrong state. Nobody asked me to audit a government PDF. Reconciling is simply what I do to data before I trust it.
An engineer would have shipped those coordinates. An accountant would have caught them and filed a ticket. I caught them and fixed them before lunch.
The same instinct is what is useful when something is actually on fire. During the national Kronos outage I rebuilt payroll data by hand and kept a hospital operating on stopgap systems for weeks. Years earlier I oversaw 1,600+ subcontractors and executed 6,000+ site installations in 90 days — a company record — and led the national rebranding conversions of BankOne to Chase, MBNA to Bank of America, and Cingular to AT&T.
The pattern never changes. The thing has to work, the deadline is not moving, and somebody in the room has to actually know how to do every part of it. I have spent twenty-eight years being that person, usually because nobody else volunteered.
ACT IIWhat is actually broken
Section 03 — the month-end close, and a switch in it you can throw yourself.
03 THE CLOSE
Every month, in most finance departments on earth, a group of capable people rebuild the same work by hand against a deadline and call the result a process. It is not a process. It is a habit with a due date — and it snaps the first time somebody takes leave, a source system moves a column, or two acquisitions land in the same week.
The work divides cleanly and almost nobody divides it. Repetition belongs to software. Judgment belongs to people. What actually happens is the reverse: accountants burn the first eight days retyping, and the exceptions — the part that genuinely needs a human — get whatever attention survives to four o'clock on close day, when everyone is tired and the only remaining goal is to stop.
So the design goal is not a faster close. It is a smaller one. Take the repetition out entirely, and hand the humans a queue of things that are actually interesting: the anomaly, the missing activity, the distribution that does not belong, the number that balances beautifully to the wrong answer.
HOW IT ACTUALLY FAILS
Nothing announces itself. One person is out. A source system renames a column. An acquisition lands mid-cycle carrying its own habits. The workbook that only ever ran on one machine opens with #REF! marching down the entire column.
Then the close does not slow down. It splits. Somebody rebuilds the file from memory. Somebody else ties out to a number that was already wrong two systems upstream. Two people chase the same variance in opposite directions and both find something. The review that was supposed to catch all of this becomes the step that gets dropped, because there are four hours left and stopping is now the objective.
Nobody is bad at their job. Every one of those failures is a design that quietly assumed nothing would ever move.
A process survives the person who built it leaving. Everything else is a habit wearing a process costume.
A conceptual reference architecture, not a diagram of any employer's systems.
Operational and source data arrives on a schedule instead of being fetched by a person.
Transactions are typed and routed. Accruals, prepaids and allocations identify themselves.
Amortization, allocation and auto-reversing logic run deterministically. Same inputs, same answer, every time.
Balance checks, duplicate detection, missing-activity detection, distribution validation. Fail loudly, fail early.
Entry, support schedule, assumptions and audit trail are produced together — because support written later is not support.
Anything the rules cannot clear goes to a human with the context already attached.
A person approves. Always. The signature is not a formality and it does not get automated.
Into the ERP, with the trail intact and reproducible from source.
Forty-seven boxes, one for every step in a month-end close done by hand. Press run and watch which of them were ever a decision.
THE CLOSE AS A HABIT
47 steps12 files6 handoffs3 hours retyping
The data arrives. Nobody fetches it.
A person reads exceptions, not formats.
The signature is the control. It stays human.
With the support and the trail already attached.
Forty-three of those steps were never decisions. They were somebody moving a number from one place to another and hoping they did not fat-finger it.
Illustrative. A generic manual close, not a measurement of any employer's process — the 47 steps are what this work looks like almost everywhere before somebody redesigns it. The point is the shape of the change, not the arithmetic.
RUN IT YOURSELF
Eight stages. Seven of them are arithmetic and routing, and a machine should do those. The one that is not is the one everybody tries to automate first. Walk it and see which is which.
Synthetic population, generic model. Not any employer's data, process or figures — the shape of the work, at the size the work actually is.
Most companies are somewhere between the second and third conversation about AI and have nothing in production. Not because the technology is not ready. Because nobody in the room understands both the model and the process well enough to say where it belongs — and, more importantly, where it does not.
The vendors will sell you a platform. The consultants will map your journey. Neither has closed your books, run your intake queue, or answered the phone when the thing broke. So you get a pilot that demos beautifully, never survives contact with a real exception, and quietly dies as shelfware nobody will admit to buying.
A model that is right ninety-four percent of the time is remarkable, and completely unacceptable for a control that has to be right every time. AI belongs in the middle, and most businesses are pointing it at the ends.
The failure mode is not the one people brace for. AI will hand you a broken map projection, a coordinate in the wrong hemisphere, or a reconciliation that balances beautifully to the wrong number — and it will do all of it with total confidence. If you cannot tell, you ship it and you find out in an audit. Twenty-eight years of doing the work by hand is the difference between leverage and a very fast way to be wrong.
An AI you cannot audit is a liability wearing a productivity costume.
ACT IIIWhat I did about it
Sections 04 and 05 — what is already built and running, and what is on the bench now.
04 BUILT SYSTEMS
Ten files. Sole author on every one of them. Each opens into a full case study — the problem, the old process, the architecture, the controls, and what it actually changed.
Lifepoint Health · 2026
Journal entry intake for a national health system — rolled out, not yet nationwide.
Independent build · 2026
An agent that drafts the entry. A human still signs it.
Independent build · 2026 — present
Excel is the cockpit. PowerShell is the engine. The model never touches an order.
Clinch Valley Medical Center · 2014 — 2026
96+ hours a month, and a full business day off the close.
Northrop Grumman · 2008 — 2011
Written alone. In production for more than ten years.
Burch Industrial Group / KNL Holdings / KNL, Inc. · 2024 — 2025
$228K out of the annual cost base. It started with measurement, not cuts.
Independent build · 2025 — present
A GPS dashboard for rail-trails. Zero mapping libraries. Deployed and in use, and still being built.
Across roles · 1998 — present
Twenty-eight years of the small things that quietly ran a finance department.
OPEN CASE FILEallenhbuilds · 2026 — present
A board that runs the shop as a crew, and a plate that recomputes its own claims every time it is drawn.
allenhbuilds · 2026 — present
An AI accountant that prepares the entry, routes it for review, and is refused when it asks to post — by something outside the model.
Files awaiting release
Work I have done that is not written up here yet, either because the case study is still being drafted or because it needs sanitizing before it is public. Listed rather than omitted, because a dossier with gaps you can see is more credible than one that implies there are none.
05 THE WORKSHOP
Twelve years of accounting work left a shelf of systems behind it. I am rebuilding them clean — my architecture, a generic model, data the code generates itself — so they belong to me and can be handed to anybody. The naming rule is simple: name it after the structure that does that job somewhere you could point at from a road in Tazewell County. Nothing abstract, nothing with a lowercase i in front of it, nothing that sounds like it raised a round.
This is also my checklist. The states are honest, which is the only way a checklist is worth keeping. 3 of 7.
THE SEVEN AS ONE THING
On a railroad, an interlocking is the machinery at a junction that ties the signals and the switches to each other so that a route which would put two trains in the same place cannot be set. Not a policy against setting it. Not a review that catches it afterward. Machinery that will not move.
That is the whole argument for letting a close run unattended, and it is the only version of it I find convincing. Ballast is the substrate every one of these sits on: load the population once, convert it at the door and report what will not convert, remove rows in a fixed order, never lose one, write down why each one left, and validate after every stage rather than only at the end. The other six are the routes through it.
Interlock is what they add up to: an engine that closes a period, records a reason for every row it moved, and stops at the step that broke instead of tying at the finish and telling you nothing. A close that ties is not the same as a close that is right. The difference is the whole job.
The crushed stone under the ties. Spreads the load, stops things shifting, drains. Every line runs on it and nobody photographs it.
The part that is identical in every one of these: load a population once, convert it at the door and report anything that will not convert, take rows out in a fixed order, never lose one, write down why each one left, and validate after every stage rather than only at the end.
890,222 rows, under ten steps in a fixed order, and it ties — with the balance and every declared check re-run after each one, not once at the finish. Put the steps in the wrong order and it stops at the step that broke. Fifteen tests hold that: one takes the guard off, runs the wrong order anyway, and watches the accrual land light while every check still passes and it still ties. The benchmark that produces the row count prints the machine it ran on, because throughput with no machine attached is a number with nowhere to stand.
The structure at the head of a mine that screens coal and drops it into the cars. Keeps what is worth keeping, sends the rest off.
Scrubs an accounts payable population before it becomes an accrual. Under ten steps in an order I can defend, a reason stamped on every row that leaves, and one gap declared rather than hidden.
Runs in your browser on the demo page — synthetic data, generated and scrubbed on the page, nothing uploaded. Type a sentence and watch rows leave with the reason attached. Then move one rule out of order: the accrual comes in light and every check still passes.
The distance between two supports. Everything in between is carried.
Contract labor accrual. Same machinery as Tipple, opposite starting material: Tipple starts from what finance has and asks what does not belong, Span starts from what operations did and asks what finance has not been told about yet. Seven rules, and the accrual is what is left.
413 worked weeks in, 148 weeks of unbilled work out, and it ties. Four rates arrived as a dash rather than a number; it says so rather than pricing that work at zero.
Where cars get sorted onto the right track.
Account strings and coding. Routing a transaction to the place it actually belongs.
A weight on a string. The oldest way of checking a thing is true before you build on it.
Account validation. Checking the foundation before anything is stacked on top of it.
The railway word for the interval between two trains.
Purchase order matching lag. How long a thing sits between the moment it happened and the moment the books noticed.
A channel of moving water that washes material to separate out what you want.
Comparing two populations and removing what matches. The least glamorous job on this list and the one people ask for most.
A name you can point at is a name you can explain, and a thing you can explain is usually a thing you understood before you built it.
Totality · generated, and pointed at something realA fragment shader on a clock, not a video. The corona is six octaves of value noise domain-warped in polar space, so the streamers run radially the way real ones do, and there is nothing to loop. The stars are there because they are there — totality drops the sky about five magnitudes and they come out. The shape is not invented. A real corona is round and bristling near solar maximum and flattens into two long equatorial streamers near minimum, so the picture is drawn from NOAA's measurement of where the cycle actually is today. The arithmetic underneath is below. Take it apart, one term at a time.
ACT IVWhy you should believe any of it
Sections 06 and 07 — the record, and the one credential I would put ahead of a title.
06 CAREER RECORD
Nine chapters, four disciplines, one method. Select a chapter to open the record.
Own work · finance systems and agentic automation
Build the close software I spent twelve years wishing existed, and take on the finance-systems and automation work that keeps me inside real problems while I do it.
Scope
Tools & systems
National health system · corporate accounting and shared services
Joined a centralized accounting group that was three weeks old, and inside three months was running journal entry accounting for the whole health system — while a nationwide consolidation pulled that work out of hundreds of facilities and into one team.
Scope
Key problems
What I changed
What I built
Business impact
Leadership
Tools & systems
WHY IT ENDED
The honest reason is that I can help more from outside than in. Inside, fixing the close means one company's close, on that company's timeline, through that company's approvals. Outside, I build it once and it works for everyone with the same five days and the same five problems. I did not leave for a better title. I left because I could see the system that should exist and could not unsee it — and three hundred sites of care is where I learned exactly what it has to do.
Centralization does not create a benefit on its own. It relocates the work. The benefit comes from standardizing the shape of what arrives — and that is an accounting policy decision before it is ever a software one.
Right-of-way, vegetation management and construction
Run three companies as senior-most operator and de facto CFO, with full licensing authority and P&L ownership — while holding a full-time hospital finance role.
Scope
Key problems
What I changed
What I built
Business impact
Leadership
Tools & systems
You cannot cut a cost you cannot see. Instrument first; the expensive decisions stop being arguments and start being obvious.
Lifepoint Health · 100+ bed acute care hospital
Serve as the analytics, automation and systems resource to the CFO, Controller and department leadership — twelve years in one production hospital finance environment.
Scope
Key problems
What I changed
What I built
Business impact
Leadership
Tools & systems
Twelve years in one department is long enough to stop seeing the process and start seeing the waste. That is not a tenure problem. It is where the automation ideas come from.
Run a retail operation on numbers rather than instinct.
Scope
What I changed
What I built
Business impact
Leadership
Tools & systems
The instinct to build the tool rather than ask for one does not care what industry you are in.
Teach software engineering to college students and working professionals while working as a developer full-time.
Scope
What I built
Business impact
Leadership
Tools & systems
Teaching a thing is the fastest way to find out whether you actually understand it. Recommended, and slightly humbling.
Commonwealth of Virginia account
Build and support the applications a hundred state agencies depended on — and coordinate the response when something statewide went down.
Scope
Key problems
What I changed
What I built
Business impact
Leadership
Tools & systems
I gathered the requirements because I was the one answering the phone at three in the morning. That is still the most efficient requirements process I have ever been part of.
Execute national rollouts at a scale where coordination is the entire product.
Scope
Business impact
Leadership
Tools & systems
At six thousand sites, nothing survives that depends on somebody remembering. The system either works without heroics or it does not work.
Design national signage and architectural branding systems — and find out that the manufacturing cost is a design decision.
Scope
What I built
Business impact
Tools & systems
Where the code started. Design taught me that the expensive mistakes are made early, in the drawing, long before anyone builds anything.
RANGE
Four disciplines. One method. The connective tissue is systems thinking — whether the system is made of numbers, people, workflows, software or steel.
FINANCE & ACCOUNTING
ENGINEERING
OPERATIONS & LEADERSHIP
DESIGN & CRAFT
languages/
Python · SQL · JavaScript · C# / .NET · VBA · Visual Basic · PowerShell · Java · HTML/CSS
llm-and-agents/
Anthropic Claude API · prompt and system-instruction design · tool calling · multi-step orchestration · context management · retrieval-augmented generation · agent output evaluation · Model Context Protocol (studying the specification and integration patterns)
ai-security/
Prompt injection and jailbreaking risk · sensitive-data handling in a healthcare environment · insecure tool execution · least-privilege tool and system access · secrets management · output validation before write-back · OWASP software security principles
platforms/
ASP.NET / .NET Framework · progressive web apps · Oracle Cloud · Azure · Microsoft 365 · SharePoint · Power Platform · ServiceNow · Okta · REST/JSON APIs · Stripe · geospatial tile services
finance-systems/
Oracle Cloud ERP and Oracle Fusion · enterprise GL and ERP · hospital financial and payroll systems · Kronos · enterprise reporting environments · RPA design and deployment · dashboarding, variance and productivity analytics
practice/
SDLC · requirements defined from operational ownership rather than interview · architecture and data modeling · validation and error handling · code review · technical documentation · production deployment and long-horizon support
EDUCATION & CERTIFICATION
07 PROOF
Virginia runs on more than a hundred state agencies and offices. I wrote applications that supported critical functions across them — agency account management and statewide outage tracking — and I wrote every line of them.
Not a team. Not a vendor. Concept, data model, code, deployment, documentation and support: all mine. I gathered the requirements because I was the one answering the phone at three in the morning when something broke, so I already knew what the software had to do. Then I built it, put it into production, wrote the SME documentation that standardized the procedure statewide, and supported it.
Those systems ran for more than ten years.
Software written by one person, holding up under a hundred agencies' daily use for a decade, is the only credential I have that I would put ahead of any title.
Your move
Three instruments that only work on a phone, a file that answers questions out of itself, and four doors.
FIELD ACCESS
SEALED
0 / 3
Three physical checks, using the instruments already in your hand. Nothing is recorded, nothing leaves the device, and every check can be skipped.
FIELD READOUT
Read from the device, live, on this page only. None of it is transmitted, stored, or attached to you.
CLEARANCE GRANTED
You stood still, turned to face north, and put your phone face down on a table — to read a stranger's personnel file.
That is the same stubbornness that makes someone rebuild a close process instead of complaining about it for five years. It is most of what I am actually selling.
So skip the form. This is the direct line.
08 ASK THE DOSSIER
It answers out of this file and nothing else. No invention, and it tells you plainly when the file runs out rather than filling the gap. It is also not neutral — this is my dossier and the box argues for me, including for the parts of my record that do not fit a normal job description. That is exactly why the switch is there. Flip it and the same evidence gets turned against me, real gaps only. An advocate who cannot state the case against is a brochure.
Answers from the file, and says what the evidence qualifies him for.
09 CONTACT
FROM WHERE YOU ARE
Four doors. I am useful behind all of them, and I am available now.
THE FIRST IS A MESS
A close that takes too long, a control nobody can explain, a process held together by one spreadsheet and one person's memory. I like these. The work is diagnosis before construction, and most people skip straight to construction.
THE SECOND DOES NOT EXIST YET
Software no vendor sells because the requirement is too specific to one business. That is the work I have done most often and enjoyed most, and the part people underrate is the last step: staying until it actually works. It is also what I am now doing full time.
IF YOU ARE HIRING FOR FINANCE OR OPERATIONS
You get someone who runs the function and can also fix the system producing the numbers. Journal entry governance across 300+ sites of care, a nationwide centralization, 49 accountants — and the automation underneath, written rather than bought.
IF YOU ARE HIRING ENGINEERS OR BUILDING AI
You get an engineer who does not need requirements gathered, because he was the business owner of the process being automated. Production code since 2008, enterprise systems that lived ten years, and agents built on the Claude API against work he was personally accountable for.
IF YOUR CLOSE IS THE PROBLEM AND YOU DO NOT WANT A HEADCOUNT
Close transformation, journal entry governance, accrual automation, ERP and reporting work — scoped as an engagement rather than a hire. I diagnose it, build it, and stay until it runs without me. This is how I fund the product, so I take it seriously.
IF YOU CLOSE THE BOOKS AND WOULD RATHER NOT
I am building this full time. What I need more than anything is finance teams willing to use it early and tell me exactly where it breaks. No cost, no commitment, and you talk to the person writing the code rather than to a support queue.
THE DIRECT LINE
NOT ON THIS PAGE
It is not a secret. It is on my resume and on my LinkedIn, and I will give it to anybody who asks for it. It is simply not printed on this page, and there is a reason: the last time I put a number straight onto the internet I heard from eleven people about my car's extended warranty and nobody at all about a job.
So it is not in the HTML, the stylesheet or the content files. It lives on the server and comes back when somebody asks for it, which a scraper reading the source will never do.
Tell me who you are and what it is about. Fifteen seconds, three boxes, and the number appears right here.
THE RECORD, IN ONE PAGE
The site tells the story. These verify it. Same person, four emphases — take whichever matches the role you are filling. All four are current as of October 2026.
EXECUTIVE
Operations // Finance // Transformation
Builder of organizations and systems.
DOWNLOAD PDF · 394 KBTECHNOLOGY
Technology // Automation // Business
Business leader who actually writes the code.
DOWNLOAD PDF · 398 KBFINANCE
Accounting // Systems // Transformation
Finance leader who can fix the system underneath.
DOWNLOAD PDF · 395 KBAGENTIC AI
Agentic AI // Automation // Operations
For AI and engineering roles specifically.
DOWNLOAD PDF · 399 KBALSO AVAILABLE
VERIFICATION
Every claim on this page is meant to survive a check. Here is how to run one.
RECONCILIATION
FILE AHB-1998 · AS AT BUILD
⚠ 1 EXCEPTION · LISTED BELOW
Every count above is computed from this site’s own content files at build time, so it cannot drift from the page. Four case studies are listed as unwritten rather than quietly omitted, and every facility-level number says so where it appears. A dossier with the gaps showing is worth more than one that implies there are none — and reconciling before you trust something is, more or less, the entire argument of this file.
— EVIDENCE
A resume is a document with no error bars. Everybody knows it, so everybody discounts all of it the same amount, which means an honest man pays the same tax as a liar. This is my attempt to stop paying it. Every figure I publish is listed below with its scope, its source and its standing. The ones I cannot currently substantiate are listed too, and they are listed first.
⚠ 1 claim I cannot substantiate today
Twice corrected in one day, and this list is what caught it both times. It sat on my resume for years as “950+ industrial clients”. They were not industrial clients — they were national, local and individual partnerships. And the count is a recollection: I remember it somewhere between 905 and 950 and I cannot put my hands on the report it came from. So it is stated at 900+, which is true at either end of what I remember, and it stays on this list until I can source it. If you are going to round an unsourced number, round it the way that costs you.
Clinch Valley Medical Center · Reporting I owned at Clinch Valley — pricing, AR analysis, revenue validation and contract billing
It stays on this list until I can get back to the source. It is not used as a headline number anywhere on this site, and the reconciliation above counts it against me.
| Figure | Claim | Basis and source | Standing |
|---|---|---|---|
| 900+ | National, local and individual partnerships reported on | Clinch Valley Medical CenterReporting I owned at Clinch Valley — pricing, AR analysis, revenue validation and contract billing | unverified checked 2026-09-05 |
| 68 | Hospital campuses | Enterprise — Lifepoint HealthScope of the journal entry accounting I owned as Interim Director, Jul–Sep 2026 | verified checked 2026-09-05 |
| 300+ | Sites of care | Enterprise — Lifepoint Health68 community hospitals plus 70+ rehabilitation and behavioral health facilities and the rest of the network | verified checked 2026-09-05 |
| 49 | Accountants led | Enterprise — Lifepoint HealthThe centralized journal entry group during the national consolidation | verified checked 2026-09-05 |
| 96+ | Hrs/mo automatedThis is a facility number. It is not an enterprise number and I do not present it as one. | One hospital — Clinch Valley Medical CenterRPA deployed across reconciliation, reporting and validation workflows | scoped checked 2026-09-05 |
| 1 | Day off the closeAlso a facility number. The redesign came first — I did not automate the old process, I replaced it and then automated what was left. | One hospital — Clinch Valley Medical CenterProcess redesign and batching logic, done before anything was automated | scoped checked 2026-09-05 |
| 10+ | Yrs in production | Commonwealth of VirginiaASP.NET/SQL applications I wrote alone at Northrop Grumman, still running long after I left | verified checked 2026-09-05 |
| $228K | Annual cost removed | Three-company operating group — Burch / KNLAsset restructuring, fleet consolidation and capital planning as senior operator and de facto CFO | verified checked 2026-09-05 |
| 6,000+ | Site installations in 90 days | National Services Associates — a company record at the timeBP rebranding program, managed with 1,600+ subcontractors | verified checked 2026-09-05 |
| 2008 | Shipping production code sinceThe site carried 1998 until 2026-09-05. That is when my career starts, not when the code does — 1998 was national signage and industrial design. Corrected in four places the day I caught it. | Northrop Grumman — Commonwealth of Virginia accountASP.NET/SQL applications written alone for state agency account management and statewide outage tracking | verified checked 2026-09-05 |
| In dev | Trestle — statusThe site called this LIVE and the resume called it “concept through production” until 2026-09-05. Both read as finished and it is not. It is deployed and it works and I use it, and I am still adding to it every week. Corrected to say so. | Independent build, trestle.allenhbuilds.comDeployed, in use on my own phone, and still being built | scoped checked 2026-09-05 |
| 890,222 | Rows scrubbed, under ten steps, and it ties | Workshop — my own code, my own dataA benchmark run over a generated close population; the balance and every declared check re-run after each step, not once at the finish. The benchmark prints the machine it ran on. | verified checked 2026-09-07 |
| 4 | Real defects the answer key caught | Workshop — my own code, my own dataVerification of a generated close corpus against its own key: 418 checks, 0 failed, on both profiles. Four defects surfaced that inspection had not. | verified checked 2026-09-07 |
| ±0.5 | Worst disagreement with the published eclipse tables | Site — tools/verify_eclipses.pyThirty-four total and hybrid eclipses recomputed from the Besselian elements alone, in Python, and held against NASA's Five Millennium Catalog before the site is allowed to build. Worst disagreement half a second; nothing outside tolerance. A bad digit fails the build. | verified checked 2026-09-07 |
| 142 | Tests passing on the AI company package | allenhbuilds — the AI company package, Version 1.1The full suite run from a clean checkout of the Version 1.1 package — 24 carried from Version 1.0 unchanged, 118 written for this version. Re-run on the date below rather than quoted from the release report. | verified checked 2026-09-09 |
| 12 | Actions permanently reserved for a human | allenhbuilds — the AI company package, Version 1.1One list in the governance module, imported by the authority engine rather than retyped. Every one of the twelve is exercised through the real event path with a canary that fails the test if anything downstream is reached — including a correctly formed, fully approved request to post a journal entry. | verified checked 2026-09-09 |
| 10,000 | Candidate AI employees generated, none activated | allenhbuilds — the AI company package, Version 1.1A containment proof, not a hiring plan: ten thousand candidates generated because a rare defect shows up at volume and not at eleven. The registry's SHA-256 is identical before and after, every candidate is orientation-only, and the activation state is NOT_AUTHORIZED. None of them exists as a Person. | verified checked 2026-09-09 |
| 0 | Operating lanes lost when failures were injected at ten-thousand scale | allenhbuilds — the AI company package, Version 1.1Ten thousand lanes, thirty thousand messages, a thousand proposed transfers with four made to fail on purpose. Zero lanes lost, zero conversations left without an owner, zero external actions executed, and three of three recovery points restored. The original tree hashes the same before and after. | verified checked 2026-09-09 |
The fastest way to know whether I am useful to you is to describe the problem and see whether I ask better questions about it than you expected. That conversation is free, and my calendar is considerably more open than it was last week.