AI Unit Economics Sprint
A bounded ten-working-day engagement for companies with production AI workloads. Measure total cost per successful business task, test practical alternatives at your quality threshold and leave with an implementation-ready plan.
Two founding engagements available for projects starting by September 30, 2026.
$12,500 plus applicable taxes, where required. Payable upfront.
Senior support beyond the Sprint when the work needs an ongoing strategic seat or a separately scoped implementation.
Technical leadership for teams moving beyond one-model prototypes. Architecture, model-routing policy, evals, hiring, due diligence, and the boring decisions that make AI systems operate.
Project-shaped AI infrastructure work — model routing, private endpoints, specialist models, evaluation harnesses, and reproducible delivery. The same approach I publish at v37.ai, applied to your stack.
Four opinions, in case they save us a discovery call.
I write down what should happen and how I'll know it did. The spec is the contract — the implementation is the easy part once the spec is right. This is how the work survives me leaving the room.
Pinned versions, deterministic seeds, the artifact built twice with diffed outputs. If I can't produce the same thing twice in a row, the work isn't done. The recipes I publish at v37.ai are written to that bar.
I ship code with the tests that prove it works, not a deck about it. If a stakeholder needs a deck, the tests are still the ground truth — the deck is a translation.
I work on a small number of things at a time. That's how the work stays good. If I'm the right fit for what you need, I'll say so; if I'm not, I'll say that too and point you somewhere better.
v37.ai is where the recipes live. Commercial work lives here; current product exploration is Model Patchbay.
Research
Three load-bearing patches that make Qwen3.6-MoE LoRA SFT reproducible bit-for-bit on AMD MI100 / ROCm. Same inputs in, same weights out.
read on v37.ai →
Artifact
A graph-based memory layer for AI agents. Nodes hold content, edges track co-activation, weights decay or strengthen with use. The substrate behind the live graph in the v37.ai hero.
read on v37.ai →
I'm Daniele Salatti. I've spent fifteen years building data-intensive distributed systems — Amazon, Meta, The Trade Desk — and the AI infrastructure work that came out of that.
v37.io is the consulting face of the same engineer who publishes open recipes at v37.ai. When you hire v37, you get me — not an account manager, not a handoff to a junior. The same person who wrote the deterministic-stack reproducibility kit will sit in your CTO chair or write the AI infrastructure your product depends on.
That bounds what I take on. I work on a small number of engagements at a time, and the work has to be the kind I'd publish about — reproducible, opinionated, defensible. If that's what you need, the call is short.
Contact
We will confirm whether the Sprint fits your workflows, available evidence, quality threshold and security constraints. If another engagement is a better shape, I will say so.