Nearly every request we receive falls into one of these three. Read whichever matches your situation — the approach will be similar.
Described as representative situations so you can picture how we work. These are not case files from any specific client, and no outcome figures are claimed here.
Retail
A dataset that never existed
Problem
A consumer goods company wants to know how much shelf space its products occupy in traditional corner shops across the Mekong Delta. No vendor can sell that data — nobody has photographed it. Buying foreign shelf imagery is pointless: layout, lighting and merchandising all differ.
Approach
Tasks are dispatched to contributors in the exact provinces to be surveyed, constrained by GPS coordinates and a radius rule preventing repeat submissions from one spot — forcing contributors to move to a new store rather than shoot several angles in one place. As images arrive, the annotation team marks each product and price tag.
Outcome
An exclusive dataset owned by the client, sufficient to train a model that measures shelf share automatically from a phone photo — something the field sales team can run at the point of sale.
Insurance
An image archive that cannot be used
Problem
An insurer has accumulated tens of thousands of motor claims with scene photography. The data has sat there for years training nothing, because none of it is labeled. Manually annotating all of it was never going to clear a budget committee.
Approach
Images pass through AI pre-labeling to mark suspected damage regions; humans confirm part and severity instead of drawing from scratch. Cases where the AI is not confident go through peer verification. Plates and faces are masked before the image enters any step.
Outcome
The legacy archive becomes a usable training set. First-pass assessment can be automated, leaving adjusters to handle only the hard cases — the same headcount, more claims processed.
Mobility
A model trained on the wrong country
Problem
A computer vision team finds their model performs well on European data but breaks on Vietnamese roads — where motorbikes dominate traffic and signage differs from international standards. Fine-tuning on more European data changed nothing.
Approach
Street imagery collected along agreed routes, time windows and road types; vehicles and signage annotated against the client's own label taxonomy rather than a generic list.
Outcome
A dataset reflecting real operating conditions, used to fine-tune the model for Southeast Asian markets instead of forcing the original model onto a reality it has never seen.
Which shape is yours?
Describe the situation you are in. If it is none of the three above, so much the better — we like new ones.