Techtel is a product AI lab. We apply artificial intelligence to problems worth solving, building technology that moves the world forward across intelligence, software and the physical layer.
Intelligence is only as valuable as the work it can do.
The frontier of AI is no longer only what models can do in isolation. It is whether they can be built into systems that carry real work from intent to outcome: reliably, affordably and in the open world. That is the work of a product lab.
Agentic systems that take a task from intent to a finished, checkable result. We work on planning, tool use and critique loops that let many specialised models behave as one accountable system.
Products that put those systems in front of people, and the infrastructure beneath them: orchestration, rendering, evaluation, security and cost control at scale.
Early research into perception and control on embedded hardware, where latency, power and failure carry real consequences. This is where our longer-term work begins.
Our first product
turns a product into a finished film.
A team of agents writes the script, storyboards every scene and composes motion as code, so each frame is exact and every scene stays editable.
Real interfaces, real typography and the brand system, animated with precise timing rather than approximated by a video model.
Original score, sound design and studio voiceover in five languages, timed to the picture.
Research
Our research follows our products: the questions in front of us now, and the ones we expect to define applied AI.
Most agents stop at a draft. We study long-horizon pipelines in which specialised models hand work to one another and the system is accountable for the final artifact: what to plan up front, what to verify after each step, and when to stop.
When a model writes the program that produces an output, rather than the output itself, every result can be inspected, edited and rebuilt. Opzami writes its films as code. We are testing where else structured generation beats raw pixels and tokens.
Quality in creative work is hard to score. We separate what deterministic checks can verify, such as structure, timing and layout, from what needs a critic model, and we build critics that judge an output against the intent behind it.
An agentic product is only useful if it is predictable in time and cost. We work on caching, model routing, honest progress reporting and the infrastructure that keeps long-running workloads stable as they grow.
Moving models onto constrained hardware changes every trade-off. Our physical-layer work asks what perception and control look like when compute, power and connectivity are all limited.
Founders
Five founders across product, engineering, infrastructure and operations.
How we work
We start from a problem worth solving. AI is a means. We choose problems where people spend real time, money or talent on work a machine could do well, and we judge ourselves by whether that work actually gets easier.
Research and product are designed together. Systems in front of real users teach us what matters faster than any benchmark. Product keeps our research grounded; research keeps our products ahead.
No fabricated progress, no dead controls. A progress bar that invents its percentage, or an agent that reports success it did not reach, destroys trust. Our systems report what is actually happening, including when they fail.
Infrastructure quality is a priority, not an afterthought. Security, cost and reliability are designed in from the start. We would rather build something correctly once than patch shortcuts indefinitely.
The person closest to the problem owns it. We do not optimise for hierarchy or headcount. Everyone ships, everyone talks to users, and both quality and speed are expected.
