AI - A useful tool, used with care

AI is part of how we work, not what we sell. We use it where it measurably helps, keep a person accountable for every result, and build the guardrails that make it safe to rely on.

In our own work - What we actually do with it

We use AI coding agents every day, in our client work and our open source projects. These are the parts we built ourselves to make that safe and useful. All of it is public.

  • Bringing vendor Android hardware to mainline Linux. Phones and tablets ship with old, heavily patched vendor kernels. On the Lenovo Tab3 7 we brought the display, the audio codec, the FM radio, clocks and resets to a current mainline kernel, shipped as a Buildroot system and as a Nerves system with A/B firmware updates, and we are porting Wi-Fi and Bluetooth step by step from the vendor sources. AI agents do much of the reading and rewriting, an engineer tests every change on the device, and each commit records that AI helped. It is the same work we do for clients like Citronics, done faster.
  • AI-assisted reverse engineering. When there is no documentation, the answers are in the vendor binaries. We pair AI agents with disassemblers and decompilers to work through stock kernels and closed userspace libraries, and turn what they find into written, register-level documentation. That is how we built an open camera driver for the same tablet, using the image tuning recovered from the stock camera library, and why our Fairphone 3 camera work is checked against reference shots from stock Android. Every finding is confirmed on real hardware before it becomes code.
  • Agents that stop where a human must decide. Nervesinator is an agentic kit that ports Elixir Nerves to Android phones and tablets. The agent does the long, tedious work: scouting the hardware, building, booting from RAM. Three steps are reserved for a person: accepting the irreversible cost of unlocking a device, choosing which partitions may be written, and deciding what gets published. Everything else is checked by scripts, not by trusting the model.
  • Guardrails that the model cannot talk its way past. Denyx puts a policy file between a coding agent and your machine. It decides which files the agent can read or write, which commands it may run, which websites it may reach and which secrets it may use without ever seeing them. Commands run in an operating system sandbox, and every action can be sent to a central audit log. It is still pre-1.0 and we say so in its README, along with the results of the automated penetration tests we ran against it.
  • Our conventions, applied by the agent. We keep a shared library of Claude Code skills that encode how we write code: naming, error handling, testing, refactoring, with idioms for Elixir, Python, TypeScript, Go, Rust and C. Agent output starts closer to what we would have written, and reviews focus on design instead of style.
  • Research, including what did not work. For three months we tested whether a programming language designed for AI could let small local models do tooling work as reliably as cloud models. Sigil matched them on single-step tasks, and fell well short on multi-step ones. We closed the project and published every number. Knowing where small models break is as useful to our clients as knowing where they shine.
  • Clear ownership. Every change is reviewed and owned by one of us. When an AI tool materially helped, the commit says so. AI is a tool we disclose, never a co-author we hide behind.

For your projects - Where we can help you use it

We start from your problem, not from the technology. Sometimes the right answer involves AI, sometimes it is a script and a database index.

  • AI-assisted development, set up properly. We help your team adopt coding agents with sensible boundaries: least-privilege access, sandboxing, skills that carry your conventions, review rules and CI checks that catch what agents tend to get wrong.
  • AI features in your product. When an LLM or a vision model genuinely improves your product, we build it with an evaluation set, measurable success criteria, cost limits and a fallback for when the model is wrong. If your data must not leave your premises, we run it on your own servers.
  • Hardware without documentation. Hardware that only comes with a vendor Android image, no datasheet and a kernel stuck years behind? We use AI-assisted reverse engineering to document it and bring it up on mainline Linux, as a Buildroot or Nerves system your team can maintain and update.
  • Traceability for regulated teams. Drawing on our fintech and compliance background, we help you document which tools are used, what data they see and who approves their output, so you can answer your auditors and your customers.

When it went wrong - Cleaning up after careless AI use

Many teams moved fast with AI and are now paying for it. We have seen these situations, and we know how to get out of them without starting over.

  • A codebase nobody fully understands. Generated code that was merged without real review: duplicated logic, missing tests, dependencies that should not be there, security holes. We audit it, tell you what is actually risky, and bring it back to something your team can maintain.
  • Agents with too much access. Coding agents or automations running with production credentials, broad tokens or unrestricted shells. We reduce their permissions to what they need, add sandboxing and audit logs, and rotate what may have leaked.
  • A demo that does not survive production. A prototype that impressed in a meeting but breaks with real users and real data. We add the tests, error handling, monitoring and evaluation it needs to become a product.
  • Costs that keep growing. Cloud model bills that scale faster than usage. We measure where the money goes, cache what can be cached, and move work to smaller or local models where they are good enough.

Our principles

  • People own the result. An engineer reviews, understands and answers for every change.
  • Measure, don’t assume. We test models on your real tasks before relying on them.
  • Least privilege. An agent gets access to what its task needs, nothing more.
  • Your data stays yours. We choose where models run based on what data they will see.
  • Honest about limits. We will tell you when AI is not the right tool.

Got a hard technical problem? Let’s talk.

A new product, a platform that needs to scale, an upcoming audit or hardware that needs great software: tell us what you are working on and one of the founders will get back to you.

Our offices

  • Cap Innove
    Rue de l‘industrie, 20
    1400, Nivelles, Belgium