Models used where they earn their keep, with limits around them
AI is a component, not a product. It is worth using where the alternative is a human doing something repetitive, and worth avoiding where a deterministic rule is cheaper and never wrong. The interesting engineering is in the boundary between the two.
On-device detection and OCR
Inference that runs on the machine rather than in a data centre: no round trip, no per-call cost, and it keeps working when the connection does not. We have shipped detection and OCR under 100 ms on a custom board.
LLM features with human review
The model proposes, a person approves, and the approval is part of the workflow rather than a policy nobody follows. Generated text that nobody reads is worse than no text.
Generated content, edited before publishing
Product descriptions and copy drafted from the spec sheet, then edited. Hundreds of items become tractable without the catalogue reading like a machine wrote it, because a person still did.
Deterministic limits around the model
Ordinary code holds the boundaries: what the model may spend, what it may change, and what a human must confirm. The model is never the last checkpoint before something irreversible.
On-device detection and recognition
Counting, sorting, or reading labels and plates on the machine itself, at video frame rate and with no cloud bill attached.
Document and receipt OCR
Scans and photos turned into fields a system can use, with the uncertain ones routed to a person rather than guessed at.
Text drafted from your own data
Product descriptions, summaries and replies generated from the spec sheet or the record, then edited by a person before anyone outside sees them.
Classification and routing
Tickets, messages and documents sorted into the right queue automatically, with the low-confidence ones handed to a human.
Research and decision support
A model reading the same inputs a person would and proposing an action, with ordinary code deciding what it is allowed to do with it.
Private, self-hosted models
Llama, Mistral or similar running on your own server, for when the data cannot leave the building or the per-call price does not work.
Catalogues too large to write by hand
Hundreds of products with no descriptions is a real cost. It is also the case where drafting plus editing beats both writing from scratch and publishing raw output.
Products that must work without a network
Retail floors, vehicles, industrial sites. If the cloud is optional, inference has to happen locally, which changes the model you can use and how it is built.
Anyone unsure whether AI belongs here at all
Sometimes the honest answer is that a rule would do it better and cheaper. We would rather say that before you commit than after.
A link beside a group means it is running in production in a project you can read about. The rest is what the team also covers.
On-device vision
In production: Edge-AI computer vision product
Self-hosted models
Around the model
In production: AI-driven trading research bot
Start with the scope conversation. It costs nothing and it is usually one call.