Last weekend, my two teammates and I spent three days at the MIT Media Lab for the ScienceClaw Hackathon 2026. We slept badly, drank too much coffee, and built something I am still proud of. We did not win.

QualityRelay at a glance: station agents, the SAFE-MESH safety gate, and the human in the loop.

Walk around the room and you would think every team was curing something. Drug discovery, enzyme design, protein folding, that kind of work. It was genuinely impressive, and it made our project look a little boring by comparison. We were in the Manufacturing and Materials track, and our project was about factory inspection. No molecules. No miracle cures. Just parts on a line, and the question of whether each one is safe to ship.

Here is the problem we picked. Take a slot and a tab that must fit together. Each part can pass its own tolerance check, and the pair can still fail to fit, because the errors stack up. A caliper, a camera, and the CAD file will also happily disagree with each other about the same part. Get these calls wrong and it gets expensive fast: unplanned downtime at a large automotive plant costs around $2.3M per hour, and the cost of poor quality typically eats 15 to 20 percent of sales.

Our 3D-printed slots and tabs, labeled by design width.

Our answer is called QualityRelay SAFE-MESH, and it works the way a good engineering team works. Instead of sending every raw camera frame to one central server that decides everything, each station owns its own evidence. A caliper agent measures the critical width. A camera agent takes a photo per part. A vision agent looks for anomalies. An assembly agent pairs the slot and the tab. They exchange small signed messages, about 4.7 kB per part instead of about 4 MB of raw frames, and a safety gate turns the evidence into one of three decisions: PASS, HOLD, or REJECT. The rule we cared most about is simple. Unsure means HOLD, never PASS. Missing, stale, duplicated, or unverifiable evidence sends the part to HOLD. And a human approves every physical action. The agents recommend; people sort and sign off.

The digital factory with the live agent panel.

The result we are proudest of is what happens when the network fails. We registered the test before writing the code, then ran our station agents against a central server on the same simulated line. On a fast link, both architectures performed identically, and we say so openly. But on a 2 Mbit/s uplink, the central server dropped to 43.2 good assemblies per hour while our station agents held 195.8. With the server down for 30 minutes, it was 89.7 against 195.8. Same safety in every case: 1.04 bad parts shipped per 1,000, against 3.90 for fixed rules. That is about 850 times less network traffic, and a line that keeps running when the factory network does not.

The key result: the line keeps running when the network does not.

We also closed the loop. In one scenario, a whole first batch comes out wrong because of a 1.6 mm tool offset error. The agents catch it, the line stops, the caliper chamber measures every suspect part, and a CNC agent computes a G-code correction that an operator approves before the machine restarts. In simulation, that run shipped 60 out of 60 good pairs with zero escapes. Feeding inspection back to a printer under drift raised good output from 154 to 195 parts per hour.

None of this stayed purely virtual. We printed eight slot and tab specimens, from comfortable clearance to deliberately defective, built a caliper chamber with an ESP32 logger that streams every reading to a CSV file, ran fit tests, and collected 88 labeled photos. Our anomaly model, trained on an NVIDIA L4 GPU, scored 0.948 AUROC on the public MVTec screw benchmark. The whole thing is backed by 550 plus automated tests, and we kept our negative results in the open: our first policy was not better than conservative rules, majority voting among agents turned out to be unsafe, and one cost-saving claim got withdrawn after we found unpriced messages. Every claim in the repo is tagged as simulated, physical capture, or external dataset, so you can check what is real and what is not.

The caliper chamber: each reading streams to a CSV row.

So why do I still believe in it, after losing to drug discovery? Because boring problems are expensive problems. Most of the world does not work in a lab curing diseases. It works in plants where a bad part reaching a customer, or a good part held for no reason, costs real money and real jobs. A system that lets a line keep deciding, safely, when the network fails is the kind of unglamorous infrastructure that factories actually need. The MIT stage rewards the spectacular. I would rather build the thing that keeps running on a Tuesday afternoon when the uplink dies.

The project is open source, and the demo runs in your browser. If you work in manufacturing, or you just like the idea of agents that say "I'm not sure, hold this part" instead of guessing, I would love to hear what you think.

Repo: https://github.com/Bakdaulet-omeraiuly/qualityrelay
Live demo: https://qualityrelay.vercel.app