What I take on
AI Cost Reduction.
Built by me, start to deploy
The same output on a much smaller bill, plus the environment and team practice that keep it down after I leave.
Why this gets built
Most large AI bills are one model doing every task, called more often than anyone measured, on prompts nobody has read in months. The spend looks like scale and is usually waste. It also rarely sits in one place: some of it is the model, some of it is the infrastructure underneath, and some of it is a vendor charging per user for something you could run yourself. The rest of it is how the work gets asked for. A team handed a subscription and left to it will re-send the same context all day, and no amount of routing fixes that on its own.
What I build
- Measurement first: what is being spent, on which call path, for which feature. A week of logs answers most of it.
- Routing, so the largest model only sees the requests that actually need it.
- Caching on the parts of a prompt that never change, which is normally most of the prompt.
- Smaller models fine-tuned or quantized for the narrow tasks, run on your own hardware where that works out cheaper.
- A harness that stops re-sending context it already sent, which is where a surprising share of the bill lives.
- The infrastructure under it looked at in the same pass, because a cheaper model on oversized machines is half a saving.
- The environment set up properly on each machine: editors, agents, MCP tools, model access, and the repository rules an agent reads before it touches anything. The same job for one person working alone is shorter, and often the one that pays back fastest.
- Prompting and context practice written for your work. What belongs in a file the agent reads once, against what belongs in every message, is the difference between a cheap session and an expensive one.
- A per-feature cost figure afterwards, so the next increase is visible before the invoice is.
How long it takes
About two weeks to a measured bill and the routing that removes the obvious half. Setting up one person on their own is about a week.
This one starts on a call rather than with a quote. Bring a month of usage and half an hour, and I will tell you where the money is going and whether I can move it. Caching and quantization run longer than the measuring does, because each change has to be proven not to alter the output before it goes near production. The environment is the short part of the rest: changing how work gets asked for takes longer, and it only holds if the team does it with me rather than watching me do it.
What ships
- A measured breakdown of current spend
- Routing layer with fallbacks
- Prompt and response caching
- Quantized or fine-tuned models where they pay
- Working environment and model access on each machine
- A prompting and context guide written for your work
- Two working sessions with the team, recorded
- Review gates, so generated work is checked before it lands
- The setup written into your repository, so somebody joining next month gets it without me
- Per-feature cost reporting
Tech stack
Next step
Let’s talk, 30 minutesNeed this built?
Tell me what you’re running and where it breaks. Half an hour is usually enough for me to say what it takes and what I’d do first.