Every product has a price the label leaves off: what it cost the planet to make, ship and bin. Ardhi reads that one. Point your camera at something, a bag of crisps or a bottle, and it scores the environmental impact across seven dimensions, then points you at a less guilty alternative. All on the phone: no cloud, no API bill, your photos stay put.
It works in steps. A vision model works out what the thing is, a second pulls maker, origin and materials, and a scoring layer turns that into the numbers.
The accuracy came from discipline, not size. Tight prompts, a scoring layer that keeps the model on a leash, and an eval harness so I could measure instead of guess. The trick was getting a 2B model to do all this offline, in a couple of seconds, without making things up.