AI Development in San Francisco, CA

AI development for an SF tech company usually means production work shipped inside their own product. SoMa and Mission startups ask for embedding pipelines indexed against their own document corpus or customer data, function-calling agents tied into their existing API surface, and evaluation harnesses that measure regression before model upgrades hit production. Buyers here have already evaluated GPT-4-class, Claude, and several open-source models on their own benchmarks, so the conversation skips the model-comparison slide and lands on infrastructure: vector database choice, retrieval strategy, prompt versioning, and observability. CCPA and CPRA shape data handling on every consumer-facing surface. The deliverable is code reviewed by engineers, not a slide deck — and the build often runs alongside the client's own AI team rather than replacing it.

RIAs, boutique wealth managers, and lending shops in the Financial District around Montgomery and California streets buy AI under SEC and FINRA marketing rules, which constrain everything a model says to a prospective or current client. Practical builds here include RAG systems indexed against the firm's research, regulatory filings, and client-communication archives, with citation-anchored responses and strict refusal patterns when output crosses into fiduciary advice. Document classifiers route inbound emails and statements. Audit-grade logging captures every model interaction for examiner review. Voice transcription and summarization for client-meeting workflows have to clear retention rules that often exceed standard SaaS defaults. The model layer is usually Claude or Azure OpenAI — both clear the compliance bar most California financial firms expect.

AI builds for Mission Bay biotech tenants — clinical-stage therapeutics companies, CROs, and UCSF-adjacent spinouts — operate under FDA promotional rules and HIPAA at the same time. Patient-recruitment chatbots have to respect 21 CFR Part 202 fair-balance and off-label restrictions, with hard-coded refusal patterns on any unapproved-indication question. RAG systems indexed against scientific literature, internal protocols, and regulatory submissions are common, and citation accuracy matters more than conversational tone — biotech reviewers care whether the model can point to a real PMID, not whether it sounds friendly. Document-processing pipelines for IND, NDA, and clinical-study submissions speed regulatory work. The model and infrastructure choices have to survive an IRB and quality review, not just an engineering one.