AI engineering is two jobs. We do both.
The first is building software with AI, where coding agents write most of the code. The second is building AI into software: assistants that professionals use during their working day. Both need what good software engineering has always needed: a way to know the result is right before anyone depends on it.
An AI engineering company is judged on what it lets through, not on what it generates.
The term is used for two different things. In the first meaning, AI engineering is software engineering done with AI: agents write code, tests and documentation, and the engineering work moves to specification, review and control. In the second meaning, it is engineering AI itself: putting a model inside a product so that a vet, a tax adviser or a teacher gets an answer they can act on, with the source attached and the risk understood.
We do both, and we found that they need the same discipline. In our method, Harness Engineering, agents write, an automated gate decides whether code may go on, and an engineer approves every merge. The same rule applies to our AI assistants: the model prepares, a person decides. The pages linked below are the method, written out in full.
- In productionYessa · our own product, for Dutch vets: a copilot that works alongside the vet during the consult.
- In productionTEX · Nextens · a fiscal assistant where every answer traces back to its source.
- In productionKai · De Rolf Groep · an AI guide for teachers inside the KlasMastr platform.
- Built forNS, HTM, P1, DJI, Pameijer, Nextens and SDU. See the work.
Agents do the routine. Engineers make the calls.
When an agent writes most of the code, the biggest change is not in the code itself, but in where engineers spend their day.
Harness Engineering
The writer and the reviewer are different agents. Code merges when the gate says so. Every mistake becomes a check with an owner.
Read the method SecurityWhat agents may touch, and what they never see.
ISO 27001, separated environments, no production data for developers or agents, and the paperwork before anyone gets access.
Read how we handle it FoundationEntry, open source.
The .NET and Angular foundation our agents build on. Your architects can check our work before we have even spoken.
Read the foundationThe model prepares. A person decides.
Three AI assistants in production, one of them in our own product. Every answer comes with its source, and the decision stays with the professional.
Yessa
Practice software for Dutch vets, with an AI assistant that works alongside the vet during each appointment.
AI in production Case · NextensTEX
A fiscal assistant where every answer traces back to the ruling it came from.
Read the case Case · De Rolf GroepKai
An AI guide that recommends courses to teachers based on their learning history and role.
Read the caseAI engineering, in plain answers.
What is AI engineering?
Two things, depending on who says it: building software with AI (coding agents write the code, engineers specify, review and control), and building AI into software (a model inside a product, with its sources and its limits engineered in). At Enigmatry both run on the same rule: the machine prepares, an automated check decides whether it may go on, an engineer decides whether it merges.
What is an AI engineering company, as opposed to a software company that uses AI?
Every software company uses AI now. The difference is whether the method is written down and checkable: which agent writes, which one reviews, what the gate tests, who signs. Ours is on this site: the method, the gates and the foundation are public, and the two ways to start have a fixed price.
Is AI-written code safe for a system that has to keep running?
On its own, no more than any other code. Behind a gate that builds, tests, scans, mutates the new tests and has a second model review it before an engineer approves it, yes, and with a fuller audit trail than hand-written code usually gets. That gate is what we sell, and what a Hardening Audit holds an existing AI-built app against.
Do we need our own data to build an AI assistant?
You need your own sources: the rulings, the manuals, the records the answer should be grounded in. You do not need to train a model. Our AI assistants use commercial models under terms that exclude your content from training, and every answer shows where it came from.
Where do we start?
With a conversation. Then either a Hardening Audit on the app you already have, or a Factory Sprint: five days to scope, prototype and design the architecture, ending in a fixed quote for a secure MVP in six weeks.
Building with AI or building AI? Let’s talk.
Perhaps both. Tell us what you are working on, and an engineer will get back to you.
Good engineering, now with AI