Sensors plus a local AI model that identifies waste on the device, with no cloud. A bin today. Drones and scanners next.
Edge AINo cloudBuilt in Europe
Artem · Aren · Mikhail · Franciszek
01 · Technology · The problem
Pollution is piling up on two fronts. Both start with not knowing what's there.
New waste, every day
Every person in the EU generates 181 kg of packaging waste a year, and only 15.3 of 35.9 kg of plastic packaging is recycled.
Old waste, hard to reach
Tourist and research litter in Antarctica. Heavy metals and chemicals in former conflict zones. Mapping is slow, manual, often dangerous.
Source: Eurostat, 2024 data.
01 · Technology · Our insight
One technology. Anywhere it's needed.
1 · SenseSensors read the item or the terrain.
→
2 · ClassifyA local AI model decides what it is, on the device.
→
3 · ActRoute it, log it, or send the right crew.
Same technology in a bin, on a drone, in a handheld scanner. Only the sensors and the model change.
01 · Technology · New places
From the bin to Antarctica.
NEXT · CONCEPT
Map from above
On a drone, a satellite or a handheld scanner, sensors chosen for the job map plastic, metal and chemical residue. Theory: not built or tested.
NEXT · CONCEPT
Send the right crew
The detected type decides who goes: metal recovery, hazardous-chemical specialists, general cleanup.
THEN
City-scale analytics
Waste data by place across a whole city.
Roadmap, not built or tested. The technology stays the same; the sensors and the model change.
Part 01 done · Part 02 now
Now: oneexample.
A smart sorting bin. SortSight, built this weekend on the Arduino UNO Q.
01 · The technology
Sense. Classify. Act.
Anywhere it is needed.
02 · The example
A smart sorting bin
The technology, working.
02 · The example · The first proof, this weekend
A smart sorting bin. The technology, working.
Camera
Looks at the item held on the trapdoor.
QRB2210 · Linux
Local model: paper/cardboard, plastic, metal, other.
STM32U585 · MCU
Turns the chute, then opens the trapdoor.
Log
Category, time and bin location per item.
LIVE DEMONo internetNo image leaves it
Sorting runs on the device. Unsure items go to Other. Known limit: clear glass and transparent plastic.
02 · The example · Source analytics · Gemma
Local AI senses. Gemma reasons.
Venue A
0
Venue B
0
0
Gemma
Every item is logged by place, category and time. Only labels and timestamps leave the bin, never images. Analytics runs on the exported log, so sorting never depends on it.
Illustrative numbers, not measured data.
Who it's for · What's next
Who it's for, what's next.
Who it's for
Cafeterias, venues and events that sort packaging waste. Cities and waste operators who need data by place.
Next steps
1 Pilot at one cafeteria or event. 2 Add a metal sensor and a larger labelled dataset. 3 Test non-camera sensors with the same pipeline on a handheld scanner, before any drone or satellite.
POSSIBLE MODEL · HYPOTHESIS Lease bins to venues, sell data by place to cities and operators.
Thank you
Sense. Classify. Act.
One technology. Cleaner sorting at the source, data by place, and a path to places no surveyor can safely reach. Edge AI is how Europe cleans up on its own terms.
Team Galilean Moons
ArtemArenMikhailFranciszek
Junction 2026 · The Green European Renaissance · Arduino UNO Q
