The AI problem your team set aside. Proven in four weeks, or you don’t pay for the proof.

Problems like these

Eight examples of the kind of problem we take on. Yours will be different.

• Transcribe and summarise meetings, with no recording or transcript ever leaving your own servers.
• Watch the camera feeds of a loading dock and flag anything out of the ordinary: a blocked bay, a skipped check, a pallet where it shouldn’t be.
• Port an application nobody maintains any more to a language your team supports.
• Score every call and chat in a contact centre against your own quality standard.
• Read documents that arrive in every layout and pull out the fields your systems need.
• Decide which claims, applications or tickets need a person, and which can go straight through.
• Check every software release against your compliance rules: access by role, audit trails, data retention. The tests repair themselves when the screens change.
• Build a pipeline that generates finished videos from your scripts, images and audio, to a standard your reviewers would publish.

How it works

1. 1

   The test week

   We agree the test, the data and the bar with you. Most teams don’t have a test yet. Helping you write one is part of the week. You keep the test set and the evaluation package no matter what. If we can see it can’t be done, we’ll tell you that week and stop. The price is fixed in writing before the build starts.

2. 2

   The build

   Four weeks. We build against the test. There are three one-hour meetings, and you get a written update each week. If we give up before the end, you pay nothing more.

3. 3

   Pass or miss

   A pass means the proof meets the bar we agreed in the test week. You pay the proof fee and use the proof in your own operations and products. On a miss, you pay nothing more. You get a written post-mortem of what we tried and why it fell short.

4. 4

   After the proof

   Hand-over is included and takes a day at most. Want us alongside your team for a couple of days while they embed it? We’ll quote a fixed price for that with the proof, and it’s remote. We don’t sell day rates.

Who carries the risk

When you buy work by the day, the risk is yours. You pay for the time, whether it works or not. With Icana Proof, we carry the risk of the build.

The catch

Yes, there is a catch and it’s straightforward.

• You always pay for the test week. You keep the test set and the evaluation package too.
• We say no. We only take on problems we expect to pass, so some stop at the test week.
• A proof isn’t a production system. It shows the problem can be solved on your data. Your team takes it into production, with us alongside if you want.

We can work this way because we’ve spent twenty-plus years learning what models can and can’t do.

Who does the work

Icana.AI’s technology leaders run Icana Proof. They set the test with you, run the build and sign off the result. The wider Icana team works with them.

Between them they’ve worked in every modality but robotics. They’ve worked at every layer of the stack, short of training a frontier model from scratch. They build and run their own evaluation harnesses and multi-agent systems every day. They’ve used every frontier model, and open-weights models too, some run locally.

Erik van Eekelen

CEO

He has a master’s degree in AI and twenty-plus years building with it. He started at Australia’s first AI company. Before Icana he was CTO, CPO and CEO in online education, logistics, e-commerce and a European smart-charging scale-up.

Recent work

• CallCoach AI, which has scored more than 7 million minutes of calls
• self-improving voice simulators for rehearsing sales calls
• RLHF systems that train models on human feedback
• autonomous marketing agents that run Icana’s own prospecting and outreach
• evaluators that pick the right open-weight model for a job

Andrew Al

Technology lead

He has spent twenty-plus years as an engineer, architect and CTO, in e-learning, carbon accounting, pharmaceuticals, agritech and marketing technology. As a CTO he was lead architect of a platform built to ISO 27001, SOC 2 Type 2 and GDPR standards. He designs the cloud, data and security platforms that Icana’s products run on.

Recent work

• fine-tuned language models for structured output
• retrieval and vector-search systems on top of language models
• data enrichment pipelines that combine generative AI, image processing and reinforcement learning
• computer vision applications, as architect and tech lead
• the MLOps automation and cloud infrastructure that keep these systems running under real load

Who it’s for

The price

An indicative A$75,000 to 150,000 in total, the test week included. Quoted per problem.

Where you land in that range depends on how hard the test is and how much data we need to work through.

Tell us the problem. If it fits, you’ll get a fit decision and a written indicative price within two business days. No meeting. The firm price is fixed before the build starts, and the A$10,000 test week counts towards it.

Your data

You share a repo or bucket you own. We work on a copy in our own environment and delete it within 30 days of the result. We settle which providers, regions and logs we use in the test week.

What you own afterwards

You use the proof in your own operations and your own products. You don’t license or resell the proof itself. We keep our tools, general methods and lessons, and we may reuse them on other problems. We never reuse your data, your labels or anything trained on them. If you want exclusivity, ask. It’s priced on top.

Describe the problem

Received. If it fits, you’ll have a written fit decision and an indicative price within two business days.