AIWAY, Edge Computer Vision Startup, Kasra Zamanian

Setup

AIWAY began as the project of four seasoned friends, veterans of the industry, taking their first real swing at AI. One of them was Giorgio, a long-time systems engineer at IBM and my mentor: the person who opened my eyes to UNIX and the Linux terminal. He was leaving, and the company needed someone young to take over as technical lead. I'd built vision systems before (during my bachelor's I made a pose-detection system that flagged non-ergonomic postures during office work), so I was the technical one. They made me lead.

The company's first product was Roadsense: a deep-learning computer-vision system that helped road-maintenance companies spot defects and anomalies on road surfaces using nothing but a phone camera mounted inside their vehicles.

Why it was hard

The model Giorgio had built was a YOLOv5 trained on a generic road-defects dataset of 30k+ images. It worked, but it wasn't tuned to Italian roads, and it had a stranger problem: it was too good. It found defects everywhere: on walls, in pedestrian areas, in parks, across banners. The catch is that they were all genuine defects. Just not on the road.

So what do you do with a model that's doing exactly the job it was trained for? Punish it and sacrifice recall? Touching YOLO was beyond my depth at the time, and fine-tuning only fixed things locally: change the city and the problem came straight back.

The model wasn't wrong; the question was

What happened: Reviewing predictions one day, I stopped trying to fix the model and asked a different question: why punish it at all? It was doing its job. It was built for detection, not localization, and localization was exactly what we needed.

The fork: Degrade the model to suppress off-road detections, or keep it intact and constrain where its detections counted.

My call: I was taking Andrew Ng's Deep Learning specialization at the time, and the segmentation lectures were the spark. Instead of weakening YOLO, let it run, then filter its detections through a segmentation model that knows road from not-road. Keep the recall, add a gate.

What it led to: A pretrained street segmentation model (MMSegmentation) worked out of the box, no fine-tuning needed. With every off-road detection (walls, bridges, parks) filtered out, precision more than doubled against the baseline. And it solved a problem we hadn't even framed yet: data privacy. Because the segmentation model already had "person" and "car" classes, we could obfuscate them and guarantee no sensitive data was stored.

Connecting the dots: detection → localization → instances

What happened: The segmentation insight opened a second product: a construction-site safety system that detected unsafe operator behavior, starting with proper use of safety gear. Same shape of problem: it didn't matter that an operator had a helmet (detection), it mattered that they were wearing it on their head (localization).

The fork: Reach for the familiar semantic-segmentation approach again, or notice where it broke down.

My call: Semantic segmentation wasn't enough anymore: with several people in frame, I had to tell my head from yours. That meant instance segmentation. By then I'd invested so much in segmentation that this was a clean extension: a straight through-line from my bachelor's pose detection, to YOLO helmet detection, to instance-level localization.

What it led to: The same architectural idea, one level more precise, and proof the lever generalized beyond the first product.

Impact

The reframe held: Roadsense kept its recall while off-road false positives effectively disappeared, and the segmentation layer gave us privacy by construction. The same approach generalized into a second product, and the team grew to four data scientists.

Why it didn't work

It worked. Until it didn't. When I joined, AIWAY wasn't a company yet; it was a dream waiting to be validated. The deal was simple: I'd build (both the technology and the team) while the others found customers and sold. We incorporated, with me as a co-founder.

But the selling never happened. We waited, the technology aged, and we kept circling the same clients who loved the demo and always needed "a little more time to think." By then we had three other data scientists, all brilliant. What we lacked wasn't technical, and it wasn't demand: the need was real across every project we touched. It was the feedback loop: glacially slow client response, and slow business development on our side.

If I did it again, I'd engage customers directly myself instead of staying behind the tech, and I'd push to pivot off the slow road-maintenance sector toward a more agile part of the market. The lesson stuck: no matter how good the product, nothing happens until you can sell it.

Takeaways

When a model "misbehaves," check whether it's the wrong tool or the wrong question. The fix is often a reframe (detection vs. localization), not a retrain.
Composing pretrained models (a segmentation gate over a detector) can beat fine-tuning, and sometimes solves problems you hadn't named yet (here, privacy).
Technical excellence doesn't sell itself. A slow sales loop will kill a great product; distribution is part of the architecture.