AI your organisation can actually use.
In a lot of places the AI would help, but nothing can answer for what it does, so the answer stays no. I bring the independent evidence and the boundaries that turn that no into a yes. An agent taking support, a copilot inside a team, a document workflow, a model on your own hardware: whatever the useful form, you put it to work knowing where it breaks and able to prove it.
Get useful AI into real work, with its own evidence.
It starts before the tooling. If your organisation wants to use AI but is not sure where it fits, or what is safe to hand it, I help you find that first: where it earns its place, what is blocking it and why, and a plan to get there. Then the hands-on work. I prove an AI system you already run, build one for a specific job, or go down to the model layer and run capable models on your own hardware. Support triage, research synthesis, an internal copilot, a document workflow, a model you host yourself: whatever the useful form, the deliverable is the thing and the reproduction that shows it holds, in your hands, not a slide about quality.
The discipline comes from building my own. This sits on Spegling, the system I build on: one memory the AI carries, a reviewer from a different family, and a ledger that answers for what the AI did. The services are the done-for-you path to the same thing, the reason an organisation that could not touch AI before can start.
The model layer is real work, not a footnote: quantization, serving on real hardware, kv-cache, keeping capable AI inside your own walls. If your problem lives down there, the writing goes deep. Read the technical work →
An AI's worst moments are the ones you never see.
Put AI in front of people or real work and you create a surface you cannot watch: thousands of unscripted moments where it can promise something false, break tone, misstate a policy, miss an escalation, or qualify the wrong lead. Pre-launch tests never saw that traffic. A latency dashboard cannot tell you it just said something untrue. So before it carries real work, I try to break it the way reality will.
Conversations you would never script
I generate thousands of adversarial cases across the ways AI actually fails in front of real work: hostile users, edge-case policy, ambiguous asks, prompt injection, the question asked sideways. Once it is live, your real traffic feeds the same harness.
A judge that does not share the blind spot
Every conversation is scored by a model from a different family than the one that produced it. An auditor that thinks like the agent misses exactly what the agent missed. That, and having no stake in the score, is what an eval you run on your own model cannot give you.
It has to hold more than once
A single good run is luck. I check the same behaviour under repetition, pass^k, so a fix that works once but quietly regresses does not slip through. This is the same gate I put on my own coding agents before a change is allowed to land.
Failure modes named, not buried
You get the failures written down and ranked by what they cost, with a reproduction you can re-run. Where it breaks is the deliverable, not the part I hide. If a job cannot be verified to a standard I would stand behind, I tell you that instead of shipping.
The proof is running on this page, not described.
The widget in the corner is an agent I built. It waits for real attention, answers only from what is written here, and refuses when a question walks off the page. It is eval-hardened with the same harness above, scored against a corpus I keep honest. Spend a few minutes, then press it and try to make it say something it should not.
Where to start.
Four are fixed scope and fixed price, no discovery deck. The fifth runs longer and is shaped in the conversation. Open one for the detail. If none quite fits, that is a conversation too: if AI could help your organisation and you are not sure how, just say so.
Adopt
Find where AI earns its place in your organisation, with the boundaries already drawn.
What's included
- Where AI helps, ranked by payoff and risk
- What is blocking it, and how to clear it
- A roadmap that ends in real use, not a deck
Prove yours
Independent evidence on an AI system you already run, so you can defend deploying it.
What's included
- Independent evidence you can hand upward
- Ranked failure modes, with examples
- The reproduction, yours to keep
Build
I build the AI for one job and hand it over proven, with its reproduction. Yours to own.
What's included
- The AI, doing one job well
- The eval harness that tested it
- Failure modes named and ranked
- A reproduction repo you can re-run
Keep it honest
Your AI kept honest as you change it, and as the models underneath it change.
What's included
- Regression caught as it appears
- A corpus that learns from real traffic
- One person who knows your system
Embedded
Some work does not fit a hand-over. I join your team for a stretch and the AI gets built inside your process, by people who will still be there afterwards.
What's included
- AI shipped inside your own process, not beside it
- Your engineers able to carry it once I stop
- Evidence built in as the work happens
- Someone who will tell you when the answer is no
The four above carry a fixed price because the work is knowable before it starts. This one is not, so the depth, the cadence and the rate come out of the first conversation and depend on your team. Say what the AI has to do and what it must never get wrong, and I will come back with a shape and a number rather than a discovery phase.
Not generic AI building.
AI ships with its evidence or it does not ship. If a job cannot be verified to a standard I would stand behind, I will say so rather than pretend.
Not a platform you log into.
Done for you, or done alongside your team. Either way you own what I build and the reproduction. Nothing to subscribe to, nothing held hostage on my servers.
One person, on purpose.
I take few engagements at a time so each gets real attention. If I am full, I will tell you, and tell you when I am not.
Not a notified body.
I produce the technical evidence that feeds your conformity and due-diligence work. I do not issue certificates, and I will not imply I can.
One conversation, then a scoped engagement.
Tell me what you want the AI to do, and what it must never get wrong. I come back with a scope and a fixed price, not a discovery call.
- The job. What the AI decides or produces.
- The line. What it must never cross.
- The traffic. Who or what it touches, and how often.
Built and verified by one person, in Helsinki. Hannu Varjoranta Varjosoft Oy · hannu@varjosoft.com