Procurement questions are narrow and they repeat: which supplier carries this product, at what unit price, in what quantity, when did we last buy it.
Answering them does not need a frontier model behind an API. It needs a small model that knows the catalogue — and, more to the point, one that runs on a machine inside the company, because supplier pricing is exactly the kind of data that is not allowed to leave it.
That last constraint is usually what decides whether an AI tool gets used at all, and it is the one an API cannot satisfy.
What it produces
A pipeline that takes real procurement records and ends with a model answering questions locally:
records → training data → fine-tune → merge → convert → run locally
The fine-tune is QLoRA on Phi-3-mini — small enough to train on a single consumer GPU. The output is converted to a format that runs on ordinary hardware and registered with a local runtime, so using it is one command on the machine it was installed on.
No API key. No per-token cost. No network call carrying supplier prices to someone else's server.
Why it is worth doing this way
QLoRA on a 3.8B model is well-worn ground; the training is not the interesting part. What matters is that the thing at the end of the pipeline is deployable by the person who needed it, rather than a checkpoint that needs an infrastructure team before anyone can ask it a question.
Open source, on GitHub.