4 posts in this category — clear
Computer vision on the factory floor: the boring parts decide it
Vision projects rarely fail on the model. They fail on lighting, mounting, and what happens when the line is dusty.
Read more
Training AI on African language data
Yoruba, Hausa and Igbo are spoken by well over a hundred million people and represented in training corpora by a rounding error. That gap is a technical problem and a commercial opening.
Read more
Retrieval-augmented generation, honestly assessed
RAG solves a real problem and introduces a new one: your answer quality is now a function of your retrieval quality, which is an information architecture problem in disguise.
Read more
Small models beat large ones for narrow tasks
For classification, extraction and routing, a small model tuned on your own data usually wins on accuracy, cost and latency at once.
Read more