Fuli VisionOn NVIDIA RTX 4090 today. On Jetson in the store next.
We already test and run vision models on an NVIDIA RTX 4090 with CUDA at Universiti Malaya, through a research collaboration. Next: NVIDIA TAO Toolkit for transfer learning on local produce, then in-store inference on NVIDIA Jetson with TensorRT and DeepStream, so each store works without a cloud round-trip.
Five stages. GPU where the work is seeing.
CPU where the work is plumbing: cameras, confirmations, reports. GPU where the work is recognising, grading and watching many shelves at once.
A loop that runs inside the store.
Public produce datasets plus the first weeks of store photos to start. Then every confirmation at the scale makes the next model better, store by store.
Each store's lighting, scale and assortment get their own adaptation, pushed to that store's device. A new store starts from the shared model and diverges from day one.
The model proposes; a person confirms. Nothing is charged or marked down without a tap. Low confidence means the normal manual flow, not a guess.
Public datasets cover the common items; local varieties will come from the founder's two stores. The first external store never sends a single image to a cloud to get started.
Numbers we will publish, per produce type.
Measured in the founder's two stores before any external store. We publish accuracy by produce type, not one headline figure, because a model that is excellent on bananas and poor on sawi is poor.
| Module | Metric | How it is measured | Reported |
|---|---|---|---|
| Scale | Top-1 accuracy by produce type | Model's first proposal against the staff confirmation, per item, per store | Per produce type, per store, monthly |
| Scale | First-tap confirmation rate | Share of weighings confirmed without a change | Per store, weekly |
| Scale | Manual fallback rate | Share of weighings below the confidence threshold | Per store, weekly |
| Grade | Agreement with the buyer | Model grade against the store buyer's grade on the same lot and photo | Per produce type |
| Grade | Markdown timing | Lots marked down on the suggested time, and whether they sold before close | Per store, weekly |
| Shelf | Empty-slot precision and recall | Alerts against a staff walk-through at fixed times | Per display |
| Receiving | Weight variance caught | Lines flagged against lines later disputed | Per supplier |
Cameras at produce, not people.
Fixed cameras look down at the scale pan and across the display. Faces are not what the models look for, and there is no customer tracking of any kind.
Store images are stored and processed in Malaysia. Recognition at the scale runs on the in-store device, without a cloud round-trip.
The training signal is the staff's own confirmation, not a third-party dataset. Stores can see and delete their images.
Compute stays on NVIDIA GPUs in Malaysia.
Today models are tested and run on an NVIDIA RTX 4090 at Universiti Malaya. Next, training on local produce will run on NVIDIA data-centre GPUs available in Malaysia, with the same rule: store images do not leave the country.
The in-store device runs inference only; training never happens on the shop floor.
