Failure-mode elimination
Repeat incidents go away on my watch. The pattern gets named, written down, and removed from the board.
About
Designer and engineer with operator instincts.
I design, ship, and operate high-stakes systems across hyperscale infrastructure, industrial operations, and production AI software.
Over 15+ years building, leading, and operating across precision manufacturing, commercial refrigeration, industrial logistics, and now hyperscale data centers and production AI software. The pattern stays the same wherever I land: take ownership of the whole system, drive repeat issues to zero, write the documentation the next person actually opens, and leave the operation calmer than I found it.
Caterpillar engine lines, Hoshizaki RMA leadership, a CNC and laser studio I founded and ran solo, FedEx high-velocity sortation, and now Microsoft Cloud Operations & Innovation. Different rooms, same instincts. Zero recordable safety incidents across full tenure in every one.
Background
Born in Griffin, built everywhere else.
Family, outdoors, and the same instinct to figure out how things work.
My name is Frank Stanton Wilson III. I was born and raised in Griffin, Georgia, and I am 34 years old. I am married with three children: 10, 9, and almost 2. Most days are a careful balance between family life, mission-critical work, and the side projects that keep me sharp.
Outside of work I am usually outside: hiking, camping, or just moving through the woods with my family. I grew up taking things apart to see how they worked, and that habit never really left. I still spend hours designing and engineering platforms and tools, partly for the problem solving and partly because building something useful is its own reward. Video games are my unwind of choice, especially anything with systems to optimize or puzzles to break.
Known For
Habits that hold under pressure.
Calm diagnosis, durable documentation, and handoffs clear enough for the next person to trust.
High-stakes calm
On cutover night, the room stays quiet. Decisions get made, the rollback is already written, and the next morning is boring.
Full lifecycle ownership
Schema to night-one logs, owned end-to-end. Leaders get one accountable person instead of a chain of handoffs.
Production-grade AI systems
AI features that pass review, behave under load, and degrade gracefully. The team can ship them without paging me at 2 a.m.
Documentation that lasts
SOPs, runbooks, and post-mortems the next person actually opens. Onboarding shrinks; tribal knowledge stops being a single point of failure.
Workflow clarity
Messy real-world complexity turned into a system the team can run on autopilot. Less firefighting, more compounding.
Throughline
How physical operations became software judgment.
Trace the fault, verify the model, document the path, and remove the repeat failure.
01
Trace → Document → Eliminate
Every failure gets walked back to its first cause, written down so the next person inherits the lesson, then the class of failure is removed. Not just the instance.
02
Read the system before the symptom
The loud alarm is rarely the real fault. Patterns in logs, schedules, vibration, query plans, and human behavior usually point to something quieter and upstream that nobody flagged.
03
Question what the room takes for granted
Most defects survive because a long-standing assumption is treated as a constraint. Naming that assumption out loud is half the fix.
04
Schema before UI
Data shape, ownership, and access rules get settled first. UI, dashboards, and reports are a downstream rendering of a sound model. The model is where flaws have to be fixed, never hidden.
05
Production from minute zero
Everything is built as if it ships tomorrow: real auth, real RLS, real failure modes, real recovery paths. Prototypes that can't survive a bad day don't teach anything useful.
Operating Principles
The runbook behind the work.
Pre-mortems, instrumentation, graceful degradation, runbooks, and alarm closure.
01
Pre-mortem before every change
Before a deploy, a cutover, or a panel swap I write down what could go wrong, what the rollback looks like, and what signal would tell me it did. The mental rehearsal catches more than the change review.
02
Instrument before you optimize
I refuse to tune what I cannot measure. Logs, traces, alarms, and dashboards land before the first optimization, so any improvement is provable and any regression is loud.
03
Write the runbook as you build
The SOP, post-mortem template, and on-call notes get drafted alongside the work and finished before launch. By go-live the next person already has the manual.
04
Design for graceful degradation
Every dependency will fail eventually. I plan the partial-failure path up front, so the system gets quieter under stress instead of louder, and the user keeps a path forward.
05
Build for the person on shift
I picture a tired operator at 2 a.m. or a candidate on a phone with one bar. If the workflow doesn't survive that, the design isn't done yet.
06
Close the loop on every alarm
No alert gets ignored, silenced, or filed without a written disposition. Either it was real and got fixed, or it was noise and the threshold got rewritten. The inbox is a contract.
A1
Structured data before generation
Models reason best over typed schemas. Free text is treated as input only, never as the contract. Inputs and outputs are validated server-side so the rest of the system can trust what came back.
A2
Retrieval before improvisation
If the answer lives in the data, the model should be looking at the data. Retrieval-aware shape on Postgres + pgvector keeps the context grounded instead of generic.
A3
Evaluation before scale
Every AI surface saves its revisions and decisions so quality regressions are visible across versions. No silent drift between a demo and production.
A4
Human fallback by design
Agents pause and hand control back when confidence is low. The user always sees a clearly labeled manual path forward.
A5
Observability is part of the product
Token usage, latency, tool calls, retries, and errors are first-class telemetry. If I cannot see it, I cannot operate it.
A6
Security at the data layer
Row-level security, scoped tools, and identity isolation live in Postgres and the gateway, never in the prompt body. The model never sees data it shouldn't act on.