Marek Gieroń

Founder. Turning factory & warehouse cameras into real-time AI sensors for safety, quality & throughput. Real deployments.

[Marek Gieroń @GieronMarek]

@JohnSmith69m @CodyaAI @roboflow Schade, dass du das so siehst. Fördergeld ist hier keines geflossen - das Projekt läuft intern und wird aus eigener Tasche bezahlt.
Wir sind eine kleine Gruppe von Technikbegeisterten, die menschliche Fehler mit KI reduzieren wollen. Mehr steckt nicht dahinter.

[Marek Gieroń @GieronMarek]

The Danger Zone prototype from my team at @CodyaAI : a drone over a park, a moving camera, cones placed one by one. A fine-tuned RF-DETR model from @roboflow finds the cones, the no-go zone is built on top.

[Marek Gieroń @GieronMarek]

@Hkempfert @CodyaAI Thanks! To integrate with a surveillance setup, we just need access to the CCTV system or a specific camera. This allows us to process and analyze the video stream in real-time.

[Marek Gieroń @GieronMarek]

🟩 = hard hat on. 🟥 = violation.

This is real-time PPE detection built by my team at @CodyaAI. It flags safety violations the moment they happen.

[Marek Gieroń @GieronMarek]

Danger Zone, part two.

Last time the zone was drawn by hand. Now it's built from the cones the crew puts down - three minimum, each within a set distance of the next.

The lone cone on the left is detected and ignored.

Prototype from our team at @CodyaAI

[Marek Gieroń @GieronMarek]

Danger Zone, part one.

A ground worker enters the restricted area around the engine, the model flags it.

Prototype from our team at @CodyaAI

[Marek Gieroń @GieronMarek]

Watch what the model ignores.

A vest and a helmet on a hook - no detections. I put them both on - detected.

Detecting the gear is straightforward computer vision. Deciding whether someone is actually wearing it is business logic on top.

Prototype from our team at @CodyaAI

[Marek Gieroń @GieronMarek]

New point of view - this one's from a drone.

Vest detection on a construction site.

Prototype from our team at @CodyaAI

[Marek Gieroń @GieronMarek]

"There's no smoke without fire"

We retrained our smoke model to catch flame too.

Fire isn't always there, and smoke fades into the background. Either signal alone raises the alert.

Prototype from our team at @CodyaAI

[Marek Gieroń @GieronMarek]

Unlike a smoke detector, a camera doesn't have to wait for smoke to reach it.

Everything we detected before - capsules, tiles, bolts - had edges. Smoke doesn't. No fixed boundary, a new shape every frame, fading out as it spreads.

Prototype from our team at @CodyaAI

[Marek Gieroń @GieronMarek]

https://t.co/HBFaw936id

[Marek Gieroń @GieronMarek]

AI that looks over your shoulder 👀

The model tracks every bolt and tells inserted from fastened, live.

Prototype from our team at @CodyaAI

[Marek Gieroń @GieronMarek]

We trained a model to find crack.

Roof tiles on a conveyor, hairline cracks flagged as they pass, with a running count of production and defects.

Model: RF-DETR (@roboflow).

Prototype from our team at @CodyaAI

[Marek Gieroń @GieronMarek]

Defects on ceramic tiles, caught on a moving line.

Fine-tuned RF-DETR from @roboflow. Fixed camera, consistent lighting, identical objects - the conditions do half the work. Change them and you're back to retraining.

Prototype from our team at @CodyaAI .

[Marek Gieroń @GieronMarek]

Detecting object size from video - possible?

Turns out yes, and simpler than expected:

Fine-tuned RF-DETR (@roboflow) on a custom package dataset → took each detection's area → measured its share of the frame → mapped three ranges to SMALL / MEDIUM / LARGE.

Then counted, labeled and wired it into a live stream.

Prototype by the team at @CodyaAI .

#ComputerVision #AI #VideoAnalytics #MachineVision

[Marek Gieroń @GieronMarek]

@omglolbah @roboflow Cool man, glad you got it working 😉

[Marek Gieroń @GieronMarek]

Counting 100 capsules by hand? No thanks.

We trained a model to count capsules on a conveyor in real time - hundreds of identical objects that touch and overlap.

Under the hood: fine-tuned RF-DETR (@roboflow).

[Marek Gieroń @GieronMarek]

Counting 100 capsules by hand? No thanks.

We trained a model to count capsules on a conveyor in real time - hundreds of identical objects that touch and overlap.

Under the hood: fine-tuned RF-DETR (@roboflow).

[Marek Gieroń @GieronMarek]

Small. Medium. Large.
With my team from @CodyaAI, we trained an RF-DETR model from @roboflow to recognize package size in real time.
#ComputerVision #WarehouseAutomation #IndustrialAutomation

[Marek Gieroń @GieronMarek]

🟩 Labeled. 🟥 Missing label.
Real-time shipping label detection, built by my team at @CodyaAI. It watches the conveyor and flags every unlabeled parcel.

Under the hood: we built it on RF-DETR from @roboflow

[Marek Gieroń @GieronMarek]

@ooooooookke Dzięki! Dotrenowaliśmy model open source RF-DETR od @roboflow na własne potrzeby.

[Marek Gieroń @GieronMarek]

Some say 7 seconds is short... I say it's enough😉

The 7-second video below is the simplest and shortest demo I've ever made! And yet, it's the one I'm most proud of😎