# AI consulting and development in San Francisco, CA > The Bay Area's software industry, from SoMa's AI labs down the 101 to San Jose, needs no introduction, and much of its AI work is now maintenance. If you built a system quickly and now have to make it durable, governed and affordable, that's the work we do here. Canonical: https://www.techfabric.com/ai-development-san-francisco Serving: San Francisco, Oakland, San Jose, Palo Alto, Berkeley, Mountain View Headquarters: 1530 E Williams Field Rd, Ste 201, Gilbert, AZ 85295. No office in San Francisco. --- ## Why San Francisco work is our work Startups in SoMa and Mission Bay shipped AI early, and many now run an agent that loses context on deploy, retrieval nobody can audit and a model bill nobody forecast, with no engineers spare to fix it. Our senior engineers build the durable half, with agent loops on a runtime that survives a deploy, RAG that cites what it read, evals that attach a number to every change, and cost instrumented from week one. ## AI consulting in San Francisco Most of our AI consulting in San Francisco is for SoMa and Peninsula startups that built fast and now need the system to hold up. We audit the agents and retrieval already in production, measure accuracy and cost, and give you a prioritised list of what to fix, govern or rebuild. [AI consulting in Phoenix](https://www.techfabric.com/ai-consulting-phoenix) ## Work a buyer here would recognise Canvass is a production Cloudflare application for account research and campaign review. Campaign preparation, approval and delivery states are kept distinct, and per-tenant isolation is enforced in the data layer, which is the study to read if your own system now has to show who approved what. ## The bill arrives before the governance does A lot of San Francisco's AI systems went from demo to production inside a quarter, on a per-token model API and a vector database nobody budgeted. The bill shows up first. Then an enterprise customer's procurement team wants an audit trail, and from January 2027 the California Privacy Protection Agency's 2025 rules on automated decision-making technology will give consumers notice and opt-out rights when such a system makes a significant decision about them. ## Questions ### We built this fast and now the model bill is a problem. That is the most common brief we get here. Model spend, retrieval infrastructure and the human review a system still needs get instrumented, and then the architecture questions become answerable. Usually the fix is durability and caching rather than a different model. ### Do you have an office in San Francisco? The office is in Gilbert, Arizona, with teams in Amsterdam, Dnipro and Hyderabad. Delivery is remote or hybrid, and we travel for discovery, workshops and the opening weeks of a build. ## What ai consulting and development covers ### Agents that hold state A task that runs for three days, calls six systems and waits twice on a person cannot live in a request handler. We build agents on a durable runtime, so a deploy in the middle of a run does not lose the run, and a retry does not send the same email twice. ### Retrieval that cites what it used An answer nobody can trace is an answer nobody can defend. Retrieval is built so every response carries the documents behind it. ### Evaluation, before and after Without a score, every prompt change is a matter of opinion. We put deterministic scorers and model-based judges in front of the thing so a change ships with a number attached rather than with a feeling. ### Human approval where it belongs Some actions should stop and wait for a person. That gate is part of the system rather than a policy document, so an agent that wants to issue a refund pauses, the approval is recorded, and the run picks up where it left off. ### Cost you can see Model spend, retrieval infrastructure and the review time a system still needs are the numbers that decide whether it stays live after the first quarter. They get instrumented at build time, not discovered on an invoice. ### Governed data access An agent reads through the same grants a person would. On Databricks that is Unity Catalog, so a model that tries to reach a table it has no grant for fails the way any principal fails, rather than because the prompt asked it nicely. ## Working with us from Arizona Arizona does not observe daylight saving, so from March to November we are on the same clock as you and from November to March we are an hour ahead. Either way the working day overlaps almost completely. Most delivery is remote or hybrid. We travel for discovery, workshops and the opening weeks of a build.