AI agents your operators can audit.
I build AI agents that connect the systems your team already uses and take over the manual work between them. Every consequential decision still goes to a person.
Start a projectApproved
approved · logged (sample)
I've done the plant work myself.
Inspection programs with plant leadership across 4 copper-mine facilities. Est. $30M+ in potential downtime avoided.
I put the human gate where failure costs most.
The consequence ranking behind FMEA and RBI, applied to what an agent may do.
I define “correct” before I build.
Formerly Boeing systems engineering: I owned product-line requirements and worked with the teams that verified them.
My tools show their work.
They either cite a source or say how far to trust the answer.
Method
Four stops on the same drawing. Each one leaves behind something you can read and check.
Diagnose
I start with two weeks inside one workflow, working with the people who run it. I map where an agent would save time, where a person has to sign off, and how I'd test it. It ends in a go or no-go, not a pitch for a bigger build.
Opportunity map, row 03Sample - workflow
- PM work-order planning
- today
- planner drafts each plan by hand from FMEA and failure history
- risk
- wrong task scheduled on a critical asset
- gate
- planner approves before the CMMS commit
- eval
- held-out set of past work orders
- verdict
- GO
Build
Then I build the agent into the systems the work already lives in: the CMMS, the document store, the failure history. It reads widely and writes narrowly. Every write it can make is listed, and every consequential one goes through a gate.
Agent tool manifestSample - read
- cmms.work_orders
- read
- fmea.library
- read
- history.failures
- write
- cmms.plan_draft, gated at ③
Govern
Every consequential action stops at a human gate. The reviewer sees the proposed change and the sources behind it, then approves or rejects it. The decision goes into an audit log with a name and a timestamp.
Approval log entrySample - action
- commit maintenance plan to CMMS
- proposed by
- agent, with 3 cited sources
- reviewed by
- planner on shift
- decision
- APPROVED
Prove
Before anything reaches production, the agent runs against a held-out set of cases it never saw during the build. Under the Operate retainer the evals keep running after launch and I monitor for drift.
Eval report, held-out setSample - cases
- past work orders, unseen during build
- scored on
- plan matches the planner's; citations resolve
- release
- held until the agreed threshold is met
Selected work
Detail sheets from recent projects. Each one states its status plainly: pilot, MVP, estimate or in progress.
W-203Est. $30M+ in potential downtime avoided
Copper-mine reliability program
Risk-based inspection, FMEA and RCM programs across four facilities at one of the world's largest open-pit copper mines.
RBI · FMEA · RCM · SAP · GE APM · Python
W-204Pilot
Excalibur
A browser-based power dialer for a national insurance organization, replacing a legacy call system. Currently a pilot.
W-201In development; method and generator shown
Corro
A method and a synthetic drawing generator for turning piping and instrumentation diagrams (P&IDs) into a reviewable equipment-and-connectivity graph.
Python · pycairo · NumPy · Pillow · pytest
Index
| Sheet | Title | Status |
|---|---|---|
| W-202 | BESSie | Internal library |
| W-205 | CogKeystone | Scoring engine built; platform in progress |
| W-206 | Meridian | MVP, pitched to a VC firm |
| W-207 | codegraph / code-council | In active development |
Engagements
Each engagement has a defined length and a defined set of deliverables.
Diagnostic sprint
About 2 weeks, fixed scope
One workflow, studied closely, ending in a clear recommendation.
Deliverables
- Opportunity map
- Gate design
- Eval plan
- Go / no-go
Agent build
About 8 to 12 weeks
The agent from the sprint, built into your systems and proven against held-out cases.
Deliverables
- Production agent
- Integrations
- Approval gates
- Eval suite
Operate retainer
Monthly
Ongoing evals and fixes after launch.
Deliverables
- Monthly evals
- Drift monitoring
- Iteration
Testimonials
“Austin took the time to learn how agents actually work the phones and put proposals together before he wrote any code. The tools he built fit the way we sell instead of forcing a new process on us, and when something needed to change, he turned it around fast.”
ApprovedMiller ComoInsurance Agent
About the engineer
I studied aerospace engineering at the University of Cincinnati (BS, 2022). There I led the propellant-processing subsystem on a NASA RASC-AL team that won Most Innovative Technology.
At Pinnacle I implemented risk-based inspection, FMEA and RCM programs across four copper-mine facilities. The result: est. $30M+ in potential downtime avoided through the redesigned maintenance program. From 2023 to 2026 I was a systems engineer at Boeing, owning the product-line requirements in a model-based systems engineering workflow.
In both places, consequential changes get a sign-off and nothing ships until it has been verified. AB Software applies the same discipline to AI agents. It's a one-person practice, so the engineer you talk to is the one who builds it.
Based in Seattle, WA. Previously held DoD Secret clearance.
Who I don't work with
- Teams that want an agent to take consequential actions with no human sign-off.
- Projects that need to reach production before there's a way to test them.
- Anyone looking for an AI strategy deck rather than a working system.
PHOTO: pending
| Rev | Date | Description |
|---|---|---|
| A | 2022 | NASA RASC-AL: Most Innovative Technology |
| B | 2023 | Pinnacle, 2021 to 2023: RBI, FMEA and RCM across 4 copper-mine facilities |
| C | 2023 | Boeing, 2023 to 2026: Systems Engineer II, requirements and MBSE |
| D | 2026 | AB Software: auditable AI agents for industrial teams |
Start with one workflow.
Tell me about the work, the systems it touches and who signs off today. Then book a 15-minute intro call.
Start a projectOr email austin@absw.io