What we believe
Software is there to do a job. If it does not make something meaningfully better for a customer, an operation, or the business, it is just more software to own.
Software design is a social, interpersonal process. It takes people, reflection, and collaboration to build great systems. AI cuts the effort needed to write code, which puts even more weight on human-led engineering: choosing well, checking the work, and shipping without drama.
More code is not the point. Useful change is.
What AI actually changes
Code is disposable. Specifications are durable. When code is cheap to regenerate, the reviewed specification—what the system must do and how it is designed—is the only asset worth keeping.
Judgment is the premium. Well-designed products come from people reflecting, debating trade-offs, and collaborating on what matters. Less time typing code means more time deciding, directing, and asking if the evidence is good enough for production.
The distance to the solution shrinks. Domain experts can now get closer to the work. Give them a safe lane, and the gap between a business problem and a working system disappears.
Chat history is not an audit trail. Mission-critical systems need decisions and evidence that survive the tool, the model, and the conversation that produced them.
Constraint unlocks smaller models. Formal specifications and our WebAssembly runtime already cut token spend and let smaller, cheaper models handle bounded slices. Next, as our build adapter pre-scaffolds crates, types and scenario tests, the coding step shrinks enough for local, open-weight models: early proof-of-concept runs already build Wasm components with untuned qwen3-coder-next — pointing toward client-hosted generation with no cloud API dependency.
The path becomes a loop. Next, production signals are investigated and fixes proposed—then specified, built and merged like any other change. Nothing bypasses the path.
Why clients trust us with it
We have shipped the hard stuff
Transport, payments, public services, real-time operations: software that is not finished when the demo works. We have spent more than two decades dealing with everything that comes after.
Senior people stay close
The people making product and technical calls work directly with the people building and checking the result. Important decisions do not disappear into hand-offs.
The controls travel with the work
Shared engineering controls put agreed tools, data boundaries and review gates into editor configuration, agent instructions and pipeline checks. The rules sit where the work happens, not in a policy document beside it.
The record outlives the agent
Our spec-driven development engine keeps intent, evidence and checks outside disposable chat history. Swap the model; the record stays.
We do not start from an empty repository
We built our own WebAssembly runtime — in Rust, on Wasmtime — to give our engineers a safe, consistent place to start. It now gives coding agents less technology to invent and reviewers less code to second-guess.
We hold generated code to the same review and verification standard as hand-written code. The useful result is not one plausible output but a delivery system that turns failures into reusable checks, guidance and tooling. See the evidence and its limits.
Both the engine and the runtime are open source under MIT/Apache-2.0 (augentic/emery, augentic/omnia). Read the code. Don’t take our word for it.
How we behave
Start with the job. What needs to change, for whom, and how will anyone know?
Say when we do not know. An honest gap is safer than a convincing guess.
Put something real in production. Small enough to control. Useful enough to teach us something.
Reuse the boring bits. Spend attention on the client’s problem, not another bespoke platform.
Leave the keys. Clients keep the code, specifications, tests, infrastructure, pipelines, evidence and decisions.
Where we work
Propellerhead is based in Auckland and works with organisations across New Zealand and Australia. We are at our best in environments such as government, financial services, insurance, utilities and health, where both the outcome and the consequences of unsafe change are material.
Start with one real change. Make the next one easier.
Bring us an AI experiment that needs a safe path, a useful change that is taking too long, or a system everyone is afraid to touch. The first conversation is with people who understand delivery.
Start a conversation
