AI systems engineering · any industry
AI changes every week. Knowing where it goes doesn't.
Every few months AI can take on another piece of the work. Someone has to decide which piece, put a check on it, and move it back when it slips. That's the job I do. I've done it for thirty years with people and software, across aerospace, legal services and healthcare. AI is one more kind of worker in the same design.
Where each step goes
Every piece of work is a string of steps, and every step belongs in one of three places. Get it wrong and AI is either wasted or trusted where it shouldn't be. Get it right and your best people spend their day on the part only they can do.
The judgment
The gray zone the rules can't settle. Anything that leaves the building, moves money past a limit, or needs a name on it.
Three kinds of work
Reading text into fields, making a call with its confidence attached, and drafting words for a person to send. Nothing else.
Everything code can do
Looking things up, amounts, dates, matching, moving a record. If code can do a step, code does it — the same every time, never a guess.
See a map move. A made-up wholesaler's bill paying, through a year of better AI: steps move to a clerk when their record earns it, one moves back when it slips, and some never move at all.
Open the work mapWhy this part doesn't go out of date
The tools are the easy part. I'll teach your team the tools in three mornings, and next year they'll be different. That's fine, because the tools were never the hard part.
Better AI makes it bigger
Each improvement means another step could move. Someone has to decide which one, on what evidence, and what checks it afterwards.
Moves need a record
A step moves when its record clears a bar agreed in advance, and moves back by itself when the record slips. Not on a demo, not on a hunch.
Some steps never move
Approvals over a limit, changes to where money goes, anything sent under your name. Deciding which those are is a design decision, made once and kept.
It's an old problem
People were always the unpredictable part of a business system. Good systems gave judgment a small, recorded place and let code do the rest. AI fits the same design.
Thirty years, across industries
Business systems in regulated work, where a mistake gets audited. Each industry change meant learning a new business fast, from the people who know it best, and finding where the work really happens.
Aerospace manufacturing
FAA · ITAR · ISO 9001
ERP conversions, plant systems kept running, and the reports management had been missing.
Legal services
E-discovery · SAS 70 · DOJ production specs
Round-the-clock document processing across teams, with a quality check at every handoff between them.
Healthcare
HIPAA · HL7
Billing systems, and a way of connecting each new medical record system that a help-desk person could run in an afternoon.
The same design, before AI
From a medical billing system I built long before AI. The clerk was the unpredictable part. Swap the clerk for an AI and every rule still holds.
- Judgment only in the gray zone. A person could approve a claim the system was unsure about.
- Fail closed. Nobody could submit a claim the system knew would be denied.
- Limits with a tripwire. The front desk could adjust a price within limits; too many adjustments emailed the manager.
Freed time was never a reason to cut people. It went to the work that wasn't getting done, and the business could do more without adding cost. And moving between industries is the skill that matters most now, because almost everything is moving.
About the name. Nova means new, syn means together. AI is a new way for people and computers to work together.
Two books, free
One for the person who decides, one for the person who builds it. Read them here or take the EPUB — no sign-up.
The AI Multiplication Doctrine
Cut the people AI makes faster, or make each of them ten times more capable? The case for multiplying, with the economics. 18 chapters, an evening or two.
Cognition Systems Engineering
Where each step goes, written out in full: which work goes to a person, a model or plain code, the checks between them, and the evidence for moving a step. A graduate textbook, 27 chapters.
Proof you can open
Working systems you can use right now, no sign-up. Each shows one idea, and they plug into each other when a job needs more than one.
Agent Desk
Make AI agents and give each one a job the system enforces. They work on a private copy of a small bakery — a double charge, a nut-allergy question, a café order to cancel — and you see every step: what they looked up, what they changed, and what the rules stopped.
Open Agent Desk- Describe the job in plain words. A small model trained for one decision suggests the permissions — anything that moves money is left for you to decide.
- Set limits. Refunds up to an amount; orders moved only so far. Checked on every action, not just requested.
- Give it work, then review it. Approve or reject each job. Turn on the demo switch and watch the system stop an agent that reaches past its limit.
Work map
One process laid out step by step — person, AI clerk or system — with the check at each handoff. Step through a year and watch it change.
Shows: the design that decides whether AI helps. Liveregistry.novasynai.comAgent Registry
Signed, versioned agents. Build one in Agent Desk, publish it, and install it into another app by address — the signature is checked first, and the new home holds it to the same limits.
Shows: rolling agents out like software, with provenance. Livebuild.novasynai.comBuilder
Describe an app, get an app. The AI writes it in a small, strict language, every line is checked before it runs, and it runs in your browser.
Shows: AI used at build time, so what runs is predictable. Livenovasupersyn.com/officeOffice
The pieces put together: a back office — customers, quotes, invoices, books — with AI clerks in it, each working for one of your people, a review inbox, and a sandbox company with six months of history.
Shows: what a whole build looks like, end to end.Document review
A small, fast decision model reads a large public email archive and flags what matters — each call explained, every one on the record, a person deciding what counts.
Shows: a model you can own, doing one judgment at scale.What every build has in common
The controls your auditors, IT and compliance people already expect for people — roles, approval limits, sign-off, an audit trail — applied to the AI. And the people who know the work are the ones who teach it.
A job the system enforces
Each AI gets the tools its role needs and limits on how far it can go. Anything outside that is refused before it happens — and the refusal is on the record.
Drafts, not sends
Anything that leaves the building — an email, a post, a refund over the line — waits for a person. Nothing goes out under your name unread.
Everything on the record
What was looked up, what changed, what was stopped, which model, how long it took. You can always see why.
Models you can own
Open-weight models throughout, and small purpose-trained ones for quick decisions — models you could run yourself, not a rented black box.
Your people teach it
The people who do the work write down how it's done. It's tested on past work before anyone relies on it, and whoever taught it gets the credit.
Small pieces that connect
Each piece works alone and talks to the others over one simple protocol. Start with the one that saves you the most time; add the next when it earns its place.
Working together
One engagement: three mornings, Tuesday to Thursday. Your team does the work with me, so they can run it without me.
Tell me about the work
What eats your week, where it lives today, and what you'd never want a machine to decide. We pick one team and one piece of work, and the first morning maps it.
Three mornings with your team
We deploy it to your own account and set it up for that work. Your people use it alone each afternoon and bring their questions back the next morning.
It's yours
The running system, the source code, and the written guidance for maintaining and changing it. No subscription and no support contract. If you want me to keep deciding what moves next, I keep a few standing mornings.
A good fit: work that happens on computers — enquiries, orders, tickets, bills, documents — and a team that wants the routine off its plate while keeping the decisions. Not a fit: AI acting on its own with no one checking.
Run a single plant? Our NovaSynAI line is built for owner-run manufacturers: a few hours a week, remote, month to month.
See NovaSynAITalk about yours
Tell me what your business does and what you'd hand off first. A person reads every message.
See it first
The work map takes a minute: one process, a year of better AI, and where each step ends up.
Open the work map