For work that was costing somebody their afternoon
AI integration
We have shipped AI into three production systems. All three do one narrow, checkable job, which is the only kind we take on.
- Jonga OCRLive
Photograph a South African ID, licence or passport. Get back structured JSON your code can use.
- RFQ ShieldIn build
Forged government quotation requests cost suppliers real stock. This reads every one that lands and says which are fakes.
- The Stokvel AppIn build
Premiums, loans, interest and the year-end payout, on a ledger nobody can quietly edit.
What you get.
Document reading that admits what it cannot see
A blurry field comes back empty rather than as a confident guess. A wrong ID number is far more expensive than a blank one, and this is the difference between the two.
Sorting and triage
Classifying what arrives, routing it, and flagging what looks wrong, with the evidence attached so a person can check the call.
A person stays in the loop where it counts
Anything touching money or identity is reviewable, and the review teaches the system. Full automation is for the parts where a mistake is cheap.
Costs you can see before you commit
Per-request cost modelled up front and metered in production, so the bill is a number you chose rather than a surprise at month end.
How this one runs.
Find a job worth automating
Something specific, repetitive and checkable. If nobody can tell whether the output is right, it is not ready to be automated.
Test it on your own documents
Not a demo set. Your real paperwork, with the bad photographs and the odd formats in it, because that is what production looks like.
Build the escape hatch first
Manual entry, review queues and a clear low-confidence path exist before launch, not after the first bad week.
Watch it and feed it back
Corrections from real use flow back in, so the system gets better at your documents specifically.
Usually paired with
Questions about ai integration.
How accurate is it?
We do not publish a number, and we would be cautious of anyone who does. An accuracy figure measured on somebody else’s documents tells you nothing about yours. What we do instead is run it on your real paperwork during scoping, in front of you, and you decide from what you see.
Does our data get used to train a model?
No. We use commercial APIs under terms that exclude training on customer data, and we design so documents are processed and dropped rather than stored. Jonga OCR holds an identity document in memory for the length of the request and keeps nothing.
What happens when it gets something wrong?
It gets caught, because there is a review path for anything that matters and a confidence signal on every result. The systems are built so a wrong answer is visible and correctable, not silently written into your database.
Is this worth it for a small business?
Only where the same small task repeats often enough to add up. Reading fifty ID documents a week is worth automating. Reading two is not, and we will tell you which one you are.
Tell us what your business is running on right now.
Thirty minutes, no charge, no slide deck. We will tell you whether this is something we can build and roughly what it would take. If it is not a fit, we will say so on the call.