Straight answers.

You do, from day one — code, prompts, evaluation suites, documentation. Ownership isn't held hostage to final payment, and there's no proprietary vallesAI runtime you're locked into after we leave. This is Clause 1 of the charter we sign at the start of every engagement.

All the major ones — Anthropic, OpenAI, Google, Meta's Llama family, Mistral — plus cloud platforms like AWS Bedrock and Azure AI, and open-weight models self-hosted in your environment. We hold no reseller or referral agreement with any of them, so the recommendation comes from benchmarks on your workload, not from a partnership.

Fixed price per step. Discovery is a fixed fee; each subsequent step is priced in the previous step's written output, so you know the cost of the next commitment before making it. You can stop at any step boundary and keep everything produced so far. Ongoing partnerships are a monthly retainer you can end any month. We don't do open-ended time-and-materials.

No. We never use your data to train models for anyone else or for ourselves, and when engagements involve third-party model APIs we default to the providers' no-training, zero-retention terms — disclosed to you before any data is sent. Details are on our security & data handling page.

Discovery: one to two weeks. Prototype: two to six weeks. Production: four to twelve weeks, depending on scope. An advisory sprint or model evaluation is typically two to four weeks end to end. We'd rather quote honestly and hit the date than quote optimistically and explain later.

Yes — it's our preferred setup. We embed with your team, work in your repositories and tools, and treat handover as a continuous activity rather than a final-week event. The goal is that your team can operate and extend the system without us.

Usually, and the first step is the same as always: measurement. We build an evaluation set from your real traffic, establish where and why the current system fails, and then you get an evidence-based recommendation — repair, rebuild, or in some cases "it's actually fine, the problem is elsewhere."

Yes, routinely, including before a first substantive conversation if you prefer. Client work is confidential by default regardless — we name clients only with written permission.

Then that's the deliverable: a written finding with our reasoning, what we'd do instead, and what would need to change for AI to become worth revisiting. You pay for discovery, not for a system nobody should have built. It's Clause 5 of the charter.

We're based in San Francisco and work with clients remotely by default, with on-site time where the engagement benefits from it — discovery interviews and production launches most often.