Your best people can't be everywhere. Their judgment can.
Every business runs on expertise that took years to build — the instincts, the judgment calls, the "we just know" that separates good from great. Most AI has no access to any of it. Orbital builds AI systems grounded in how your business actually works.
I'm Johan Steenkamp, an AI product engineer in Christchurch, New Zealand. I capture that expertise, write it down, and build the system on the Document Substrate, a foundation that already exists and is in use.
Why most AI projects disappoint
Generic AI knows what's publicly true about your industry. It doesn't know what makes you good at what you do.
Your senior accountant's instinct for which deductions to flag. Your best lawyer's judgment about contract risk. The way your operations team triages exceptions before they become problems. None of that lives in a language model.
So the AI sounds plausible but flat. It gives textbook answers when your clients expect your answers. The automation handles the easy cases but misses the nuance that matters.
The fix isn't a better model. It's giving AI access to the expertise your business already has.
Intelligence is bought. Expertise is built.
The model supplies intelligence: reasoning from whatever context it is given. Anyone can buy it, and next year's model will be better.
Expertise is knowing how this firm does the job: which document to reach for, which figure does not add up, when a finding is accepted with a note and when it is chased. It lives in people and in the checklists, letters and corrections they produce. No model has it. It has to be captured, written down, and tested against real work.
What Orbital builds
Orbital turns the documents a client sends, the statements, invoices, receipts and agreements, into records a reviewer has checked and signed. The ledger owns the numbers. The general assistant owns the conversation. Orbital owns the work in between.
What I provide
- Map. Short sessions with the people who do the work, starting from the workpapers, checklists and letters they already produce and one real job followed end to end. You receive a practice manual seed in the firm's own words, a completeness specification for each job, and an architecture note. Paid for on its own, and yours whether or not we build.
- Build. Your document types, checks, workflow rules, manual and test cases, on the Document Substrate. The general parts already exist, so you pay only for what is yours.
- Improve. One install per customer, in accounts you own, managed by Orbital. Every correction a reviewer makes is recorded with its reason and becomes a test case, and we iterate on the checks, the manual and the document types as real jobs show what the first version missed.
Why build on the Document Substrate
It is deployed for a New Zealand accounting practice on its rental property, sole trader and year-end jobs, and documented end to end.
- Reads the shoebox. PDF, Word, Excel, scans and photos, arriving by upload, email or API. Document type is detected from the contents. Confident detections are applied; the rest wait for a person.
- Personal details never reach the model. Names, addresses, bank account, IRD and company numbers are replaced with placeholders before any model call and restored only for an authorised person. Every restore is logged.
- Every figure is traced to its page. A figure that cannot be found on the page is flagged, never passed.
- Checks run in code. Twenty-two document types have field definitions and eighteen have arithmetic checks. Same files, same verdict.
- The model never does arithmetic. Code adds the totals. The workpaper and draft letter are generated from checked records only.
- Quality is tested against the firm's own work. Rated answers and corrected figures become permanent test cases that gate every change. For redaction, zero leaks is the pass mark.
- The firm's judgement is written down. A practice manual in the firm's words, with a test that fails the build if the manual and the checks disagree.
How the Document Substrate works · Explained simply · Project detail
How I think
- Garbage in, garbage out. Validation checks that a file is complete and consistent before any money is spent on it.
- Doer to reviewer. Verification checks that each figure is true to its source, so people review the machine's work instead of doing it.
- Keep the model out of the parts that have to be defensible. The model reads and suggests. Code checks and decides. A person signs.
- Personal details never leave the building. Leaking customer data is the one thing a business cannot get wrong. Details are removed before the model sees anything, and restored only for the person allowed to see them.
- Ledger, not spreadsheet. Records with provenance. Excel only at export.
- One install per customer. Your documents in your own accounts, not a shared database.
- Small firms first. Practices, property managers, professional services, councils and co-ops.
- Expertise is captured, not generated. Models and prompts are commodities. Your definition of "complete" and your correction record are not.