In, out and left. The books said 42, and every system called that shelf correct. I traced the surplus to a proven wrong-lot shipment, then built the check that closes the gap.
MeasuredI design and build the AI systems that run businesses.
Architect and engineer of 42 live automations and five production systems at a precision manufacturer. They receive goods, track stock, run production and check orders before they ship.
~1,200
automation runs a day on the 30 server automations, out of 42 live.
Measured~350 hours a year
of routine staff time returned across seven automations.
Estimate~$9,000 a year
in hosted platform cost avoided by running on owned hardware.
Modeled from a list price, converted from euros~$0.15
of model spend per checked order, at 40 orders a working day.
Measured from token usage, at list ratesThe foundation
A self-hosted integration bus.
The production systems connect through a self-hosted workflow engine, with explicit webhook contracts between systems. Stations are small web apps, so updating one container updates every desk.
42live automations, 30 on the server and 12 in the cloud
9,535server runs in the 8 days measured
System 1
Dock intake
A pallet arrives with dozens of cases, and each batch has to reach the storefront and the accounting system with its own traceable record. That used to mean clipboards and retyping lot codes by hand.
How it works
- Four chained workflows behind one phone app
- A vision model reads labels, and the printed packing list stays ground truth
- Paperwork, side count and close-up count must agree before anything books
3 to 1people per pallet, working from a phone
36 of 36lines matched on the reference intake
System 2
Inventory sync and reorder routing
The production plan runs on a workbook of about 1,000 active parts. Someone used to export a report, upload it and update dates by hand each morning. Now it runs itself.
How it works
- Pulls stock live from a desktop accounting system, through an API bridge
- Three scheduled runs a day, plus an hourly poll for manual requests
- Keeps automation out of the columns people own, so a person's reassignment sticks
- Sets reorder dates and routes each part to its work station
3×a day, up from once at best
63 to 83hours a year returned, estimated
System 3
Rack-to-receipt production ledger
A stock rack that existed in no system is now tracked part by part, from raw blank to accounting receipt, across stations, travelers and labels. Its assistant can plan any floor action and cannot commit one.
How it works
- Each button press is planned in plain words, and the server enforces a person's Yes
- Single-use tokens, tied to the login, refused if a part moved since the plan
- Move and event logs are append only, enforced by the database
- Assistant evals gate each deploy: 13 cases, 5 of them safety cases that must all pass
487tests passing on version 2.0
78 of 78real floor presses matched in shadow, over 4 days
System 4
Shipping verification kiosk
Two camera stations check packed orders before they ship. A vision model reads what is in the box, and code compares it with the order system. The model never makes the decision.
How it works
- One workflow of 72 steps across 5 webhooks
- Not told the quantity or slip number, so it can't simply confirm them
- Schema-checked answers, then confidence floors by product type
- 9 conditions force a hold, and more than a dozen send an order to review
- Photos of each order stay on file as the dispute record
17wrong boxes stopped and fixed in 12 weeks, a measured floor
~9 sa verify, on one partial day of the current build
System 5
Operations watcher
A read-only console over everything. It looks for the failure dashboards miss: the system said something was wrong, and the work moved on anyway.
How it works
- Scans every 5 minutes during working hours, for known failure classes
- Reads through a read-only collector and keeps its findings in its own store
- Each rule records the real incident that created it
5 minbetween scans, on weekdays
0 writesto any system it watches
And the daily work
The routine jobs that no longer need a person.
Backorder copying, a payments check, a daily accounting report, a paid-order watch and new stock emails all run without anyone touching them.
~225hours a year across these five, estimated
30 minbetween paid-order checks
Beyond the shop floor
Adapt Automations, the company I founded.
Adapt builds websites, automation and apps for paying clients.
Lead Hunter
- Scans local businesses, then runs them through a CRM to a personalized preview site
- Previews print only facts it can trace to a source
- Outreach is blocked on the server until a person approves the preview
Nico
- A live LLM chatbot scoped to the site's own content, with a red-team suite rerun on each prompt, knowledge or model change
6,623businesses in the Lead Hunter database
2,235tests passing on the Lead Hunter CRM
The results in detail
Hours back, errors stopped, costs avoided, and problems no one had found.
17
wrong boxes stopped and fixed before they shipped, in 12 weeks, each proved by comparing the stopped and corrected photos. This is a floor. The same review counted 89 stops on boxes packed correctly.
Measured~350 hours a year
of routine staff time returned across seven automations.
Intake 42Sync up to 83Five routine jobs, about 225
Estimated by the builder, on stated minutes per task~$9,000 a year
in hosted platform cost avoided by running the workflow engine on owned hardware.
Modeled from a vendor list price, converted from euros~$0.15
of model spend per order, with each verify taking about 9 seconds at roughly 40 orders a working day.
Measured. Cost at list rates, seconds from one partial day90, 46, 44
How I ship
Each change is tested against written gates before it reaches the floor.
The release process is a system too. Gates are written before the run and don't change after it, and a person rules on any gate a change fails. Engine changes replay against stored decisions, and prompt changes run side by side against the live prompt.
12 to 69unit tests a release
13 to 37of that day's live runs
188 to 232retained runs, full inputs
290runs recounted from photos
2,830stored final decisions
Runs in this page
Try one of the kiosk's rules.
The order expects lot L26-0417-A. Write what the model claims it read, and this rule decides.
Review
Run them yourself
Five systems you can operate in your browser.
Clean-room versions on made-up data. Flip the switches, break things and watch which part catches it.
- Dock intakeRead a pallet, cross-check three sources and book lots.
- Inventory syncStep through a day of stock pulls and reorder routing.
- Production ledgerSign in, take a batch from rack to receipt, and switch to Admin to book it.
- Shipping kioskSix ways to break it, a release harness and a test of the test rig.
- Operations watcherWatch a sweep raise findings with their evidence.
Credentials
Courses and exams on the stack I build on.
| AWS Agentic AI Demonstrated, hands-on exam lab | AWS | Verify |
| Agents Course, Certificate of Excellence | Hugging Face | On request |
| Claude with the Anthropic API, Claude with Amazon Bedrock, MCP Advanced Topics, and six more courses | Anthropic Academy | On request |
| N8N101, N8N102 and N8N103, the full Foundations program | n8n Academy | Verify |
| In progress: AWS Cloud Practitioner | AWS |
Outside the code
I am an FAA commercial pilot, single and multi-engine land, with an instrument rating, and I spent six years in law enforcement. Checklists, procedure and calm failure handling were habits long before I built my first automation.
Get in touch.
Email don@doncox.ai. The code for this site and its demos is at github.com/Donald-Cox. Download my resume.