AI that debugs real hardware
We put AI agents at a physical test bench — with instrument-grade eyes and hands, hard safety interlocks, and a senior engineer in command. The result: firmware problems that used to consume weeks of bench time get found, fixed, and verified in days, with every claim backed by a hardware measurement.
The Bottleneck
The hardest problems in a hardware product live at the boundary between firmware and physics.
A simulator won't catch the charge controller that only fails when a thermistor is disconnected. A code review won't find the sensor that never sleeps and quietly drains your battery. These bugs only reveal themselves on real hardware, under live conditions — and finding them takes an engineer, a bench full of instruments, and time: flash, measure, think, adjust, repeat. That loop is slow, expensive, and it doesn't scale. It is the quiet bottleneck of almost every hardware program.
We Closed the Loop
For the past two years, Solid has been building something we believe is genuinely new: a system that puts AI agents inside the develop-test loop, not just beside it. Our agents don't stop at writing firmware. They flash it to the actual device, drive lab instruments — programmable power supplies and loads, power profilers, waveform generators, logic and protocol analyzers — and read back measurements, including continuous visual and thermal camera coverage of the board itself. Then they form a hypothesis, change one thing, and measure again. And again. Around the clock, without fatigue, with a senior electrical engineer directing the work and reviewing every result.
We run this on our own self-contained, remotely operable test cells: your device is fixtured once inside an instrumented enclosure, and from that moment the full test capability of the bench — flashing, power cycling, probing, imaging — is available on demand, from anywhere, in parallel across as many devices as the problem needs. One engineer, many benches.
Built on Hapimatic
The system behind this work is Hapimatic, our platform for running AI agents as a distributed fleet, and its Hardware Bench, which gives those agents instrument-grade eyes and hands. We built it for our own bench first. The core is open source, and the bench is available as a hosted service.
Run the core yourself
The fleet platform itself: run your agents on your own machines, reach them from any browser or phone, and keep sessions alive across reboots, network drops, and rate limits. Open source under AGPL-3.0.
github.com/SolidPD/hapimatic-core ↗ Hosted servicePut your agents at our bench
Hardware Bench, a paid add-on of the hosted service: a camera over the board, an oscilloscope, programmable supplies and loads, a power profiler, and debug probes, all within your agent's reach. In pilot, contact us.
hapimatic.com/hosted/hardware-bench ↗ Engineering servicesLet us run the loop for you
Bring us the problem. Our engineers fixture your device, direct the agents, review every result, and deliver the fix with its evidence trail. No platform to learn, no bench to build.
See engagements ↓Built for Firmware Problems
This system is ideal for solving firmware problems in hardware devices — and that's by design. Firmware iterates in seconds, its effects are physically measurable (current draw, timing, temperature, RF behavior), and the hardware itself is the objective referee: a fix either holds on the bench or it doesn't. That makes firmware the perfect material for a closed loop that can run thousands of build-flash-measure iterations while your team sleeps.
Power Optimization
Battery life is a firmware problem wearing a hardware costume. Our agents profile current draw correlated line-by-line with firmware execution, find the waste — wake locks, chatty polling, radios that never rest, peripherals left powered — then implement the fixes and verify every milliamp of improvement on the physical device.
You get back a battery-life number, and the evidence trail behind it.
Bug Identification & Troubleshooting
The intermittent fault. The protocol handshake that fails one time in fifty. The inherited codebase nobody fully understands anymore. Our agents reproduce the failure on the device itself, instrument the code, bisect the behavior, and watch the result on the instruments — logic analyzer, power profiler, thermal camera — until the root cause has nowhere left to hide.
Then they prove the fix by trying to break it.
Regression & Endurance Testing
Fixture your board once, and your full hardware test suite re-runs on demand or on a schedule — against real rails, sensors, and loads. Soak tests, characterization sweeps, and overnight captures run unattended.
The regressions simulators miss get caught before they ship.
Results We've Demonstrated
This isn't a concept page. The system has already delivered on client work:
- Power optimization, end to end. What began as an advisory review for a client's energy-harvesting device became a complete, delivered power-optimization implementation — analysis, firmware changes, and hardware-verified results — executed through our agentic system under engineer direction.
- Firmware stabilized. A client's power-management firmware was oscillating so wildly the feature was unusable. Our closed-loop process reproduced the behavior, isolated the cause, and delivered stabilized firmware verified on the client's own hardware.
- Stalled projects rescued. Inherited, undocumented designs are where this approach shines brightest: the system methodically maps what exists, tests its assumptions against the board itself, and drives the project to a working outcome. We've now done this more than once — including work our clients had written off as stuck.
- Our own lab, built the same way. The test cells themselves — instrument drivers, safety firmware, evidence pipeline — were developed with the same closed-loop process we sell.
Why You Can Trust It
AI with its hands on hardware raises fair questions. We designed for them from day one.
Enforced Below the AI
The agent never touches an instrument directly — it can only ask. A policy layer owns every instrument handle and refuses or clamps any request outside the device's safety envelope. Below that sits a non-overridable hardware safety floor: absolute thermal ceilings, smoke-triggered power cutoff, an e-stop chain, and a watchdog that fails to a safe state even if everything above it is dead.
Measurements, Not Claims
Every measurement writes a durable, append-only record: timestamped raw data, the exact intent that requested it, the device profile it ran under, a safety-state snapshot, and camera frames of the board at capture time. Refusals are logged too.
Every number in our reports traces back to who asked, what ran, and what the device looked like when it happened.
An Engineer in Command
The loop is agentic, not autonomous-and-unaccountable. A senior electrical engineer with two decades of embedded experience scopes the work, watches the same live session the AI runs, and stands behind every deliverable.
What We Can Do for You
Firmware Debug Sprint
Bring us the bug nobody can find. We fixture your device, reproduce the failure, and hunt it to root cause — then deliver the fix, the proof, and the evidence trail. Fixed scope, fast turnaround.
Power Optimization Sprint
A battery-life audit that ends in shipped firmware, not a slide deck: profiling, prioritized findings, implemented fixes, and measured, hardware-verified improvement.
Dedicated Test Cell
Your product, permanently fixtured in its own instrumented, safety-interlocked test cell. On-demand regression runs, scheduled soak testing, and a remote bench your team can reach from a browser — parallel test capacity without parallel headcount.
Something Else?
If your problem lives at the firmware/hardware boundary, it's probably a fit. Tell us what's broken — the first conversation is free.
