AI & Machine Learning Company in San Francisco

Three delivery shapes drive the range: a focused AI proof of concept built at SF rates, a production custom ML model (fraud scoring, churn prediction, document extraction, ranker), or full production AI systems with multi-step agents, fine-tuned open-weight models, evals, red-teaming, and enterprise integrations. Scope drivers include data audit, a baseline model or RAG pipeline, evals, and a hosted demo your design partners can actually use. SF rates sit above most US markets because of the OpenAI, Anthropic, and Scale talent premium, but we deliver from Edmonton and Chandigarh on PST hours, so you get Bay Area quality without Bay Area headcount cost. Fixed-fee proposals only, no open T and M.

California AB 2013, signed in September 2024 and effective January 1, 2026, requires developers of generative AI systems made available to Californians to publish a high-level summary of the datasets used to train the model. The summary covers data sources, whether the data includes personal information or copyrighted material, time periods of collection, and whether synthetic data was used. Our discovery phase runs an AB 2013 review on your training and fine-tuning pipeline, builds the public-facing summary your counsel will sign off, and structures dataset documentation so the disclosure is defensible. Note that California SB 1047, which would have imposed broader frontier model safety obligations, was vetoed by Governor Newsom in September 2024, so AB 2013 is the live obligation for now. We also flag Colorado AI Act exposure for clients selling into high-risk decision systems across state lines.

Yes. We ship RAG copilots, internal agents, and document workflows for SF SaaS unicorns and the financial and healthcare buyers they sell into. Production builds include SOC 2 Type II aligned controls, prompt injection defenses tested against the OWASP LLM Top 10, PII redaction before any data hits a frontier model, tenant-scoped logging, and eval harnesses that block deploys on regression. For HIPAA-aligned workloads we route through Azure OpenAI or AWS Bedrock under a BAA. For customers who want zero data leaving their VPC we deploy self-hosted Llama 3 or Mistral on GPUs in their AWS account. Output logs feed Langfuse, Arize, Datadog, or whichever observability stack your SRE team already runs.

The California Privacy Rights Act extended the CCPA with stronger rules on automated decision-making, sensitive personal information, and consumer rights to opt out of profiling. Our AI pipelines apply data minimization at ingestion, classify sensitive personal information against the CPRA categories, and wire the access, deletion, and opt-out endpoints required for consumer-facing AI. For B2B SaaS clients selling to California enterprises we publish the contractual language and DPA terms procurement teams expect. CPRA also interacts with the upcoming California Consumer Privacy Agency rulemaking on automated decision-making technology, so we keep an eye on the regulatory pipeline and adjust the build before draft rules become enforceable.

Most of our SF clients are either YC alumni (W23 onward, heavily weighted to the W24 and W25 AI batches) or post-YC seed and Series A teams with founders out of Stanford HAI, Berkeley BAIR, or the OpenAI, Anthropic, and Scale diaspora. When a project requires genuine novel research (rare architectures, new training recipes, novel eval methodology) we scope collaborations with HAI or BAIR affiliated researchers rather than overselling in-house science. Our core team handles applied engineering, evals, MLOps, and productionization, which is where most YC AI startups stall after Demo Day. For standard work (RAG, fine-tuning, agents on known frameworks, classical ML) no academic partner is needed, and we will tell you up front which bucket your problem fits.

We default to AWS us-west-2 (Oregon) primary with us-west-1 (N. California) secondary for latency-sensitive edges, GCP us-west1, or Azure West US 3. For HIPAA workloads we route through Azure OpenAI, AWS Bedrock, or self-hosted models under a signed BAA, keeping PHI inside the customer VPC. For EU exposure we set up parallel inference in eu-west-1 or eu-central-1 with GDPR DPAs in place. CPRA obligations are mapped during discovery and documented in the data processing addendum. We avoid sending sensitive personal information to frontier APIs without enterprise contracts that disable training on customer data, which OpenAI, Anthropic, and Google all now offer by default on their business tiers.

A focused production RAG copilot ships in eight to fourteen weeks with a team of three: staff ML engineer, full-stack engineer, and a designer if there is a user-facing surface. A multi-step agent system with tool use, eval coverage, and red-team pass takes fourteen to twenty-two weeks. A fine-tuned open-weight model with custom training data, evals, and inference infrastructure runs sixteen to twenty-six weeks. We ship to internal alpha by week four behind feature flags, so SF founders can put working AI in front of design partners and YC batchmates long before general availability. Most teams underestimate the eval and red-team portion of the timeline, which is exactly where SF buyers now apply the most pressure.

Most custom AI development qualifies for the US federal Research and Development tax credit under IRC Section 41, and California offers a parallel state R&D credit. Qualified activities include novel modeling, architecture experimentation, algorithmic uncertainty resolution, eval methodology development, and systematic experimentation, but not routine integration or off-the-shelf API wiring. We deliver contemporaneous documentation, time-tracked engineering logs, technical narratives, and failed-hypothesis records aligned with the Section 174 capitalization rules that took effect in 2022. Your CPA or specialty R&D tax firm (alliantgroup, KBKG, Source Advisors) files Form 6765 and the California FTB 3523, and our records support the claim. Final eligibility sits with your tax advisor and the IRS.