intent - AI Creating New Bottlenecks in IoT Development

This one is the peanut butter and jelly of connected hardware, a software consultancy that builds for hardware startups sitting down with a hardware development firm. Peter Tuszynski, CTO at intent, joins DeAndre Harakas and Grant Chapman to talk about what AI has actually changed about bringing an IoT product to life, and what it has not.

Peter's answer is more grounded than most. He has seen individual engineers go from 1.5x to 7x on output, but he calls that unrealized capital gains, because shipping more code just means shipping more unverified code until the whole process gets the benefit of that speed. Grant's framing matches, that the real return on AI today is de-risking rather than velocity, since teams using it well are looking under rocks they would never have turned over, and every time you speed up one stage the bottleneck reappears further down the line.

At intent the bottleneck turned out to be manual QA, because their projects involve a phone, a peripheral, and a cloud backend, and the messiest failures live between the phone and the peripheral where Bluetooth is low bandwidth and the device landscape is a mess. So they built a QA agent that is agnostic of the underlying model, wired up to drive iOS or Android, a tethered peripheral over CLI, and a test case repository, and found that newer models can read a human readable test case and work through it the way a QA engineer would.

The conversation also covers why embedded was supposed to be out of reach for AI and what actually changed, which came down to giving a model a board to flash and a way to verify what it flashed, and from there to giving models control of lab equipment and closing the loop between the physical world and what a model can touch. Peter traces the progression from prompt engineering to context engineering to harness engineering to loop engineering, and the two of them get into designing hardware to be testable in the first place, the carpenter and the CNC machine as the analogy for engineers worried about their craft, why prompting has started to look more like talking than writing, and why play, not work, is how anyone actually figures out what these tools are good for. Filip Pietraszkiewicz closes with advice for founders on getting something clickable in front of users as early as possible.