Data platforms and applied AI, built to last.

Both, and the choice is made per task. Claude or another hosted frontier API is usually right when the work needs frontier-level reasoning, the volume is spiky, or nobody wants to run a GPU. An open-weight model is usually right when data must stay inside your boundary, the volume is steady, or the cost model says so. Either way the surrounding system — the data, access controls, evaluation, and logging — is the same work, and that is most of the job.

For steady internal workloads — assistants, extraction, classification, search — usually yes once utilization is reasonable, because a well-served open model on a single GPU can process a very large number of requests. For spiky or tiny workloads it often is not. We build the cost model from your traffic before recommending anything.

Yes, and it is often the easiest place to start. We attribute spend by team and workload, fix the expensive patterns — oversized warehouses, unclustered queries, idle compute, frontier models on routine tasks — and leave budgets and alerts behind. For AI spend specifically, routing and caching usually come before self-hosting; we model which one pays off before recommending either.

That the system has been designed around what the model is allowed to see and do, that untrusted content cannot escalate into unintended actions or data leaks, that the controls are proven by an evaluation suite rather than assumed, and that prompts, retrievals, and tool calls can be reconstructed afterwards. It is engineering discipline, not a certification or a penetration test.

Whichever benchmark best on your tasks. The landscape changes monthly, so we test candidates — currently the Llama, Qwen, Mistral, Gemma, DeepSeek, and gpt-oss families — against your own examples rather than public leaderboards, and serve the winner with vLLM or llama.cpp. The selection method is what we bring; the model is a swappable part.

Dryware is the independent practice of a staff-level data engineer with more than fifteen years in backend systems, data platforms, and machine learning infrastructure — most recently leading a data team through a migration to dbt, Snowflake, and Dagster and a re-architecture onto an S3 lakehouse on Spark and Iceberg, while putting LLMs on production paths with Claude, MCP servers, and Snowflake Cortex; before that, building the model-serving and deployment system behind a machine-learning product. The person you talk to is the person who does the work.

Dryware takes on one or two engagements at a time so that each gets senior attention. Most are part time, from a half day to two days a week, which suits scoping, architecture, and build work that your team then carries forward. We say so up front when a request needs more capacity than that.

With a free 30-minute intro call. If it looks like a fit, we propose a short, paid scoping engagement that produces a concrete plan, a build sequence, and a fixed price or a day-rate estimate for the work. You know what you are buying before you commit.

Yes. Dryware is based in San Francisco and works remotely with teams across time zones, with overlap on Pacific hours. Engagement terms and data handling are agreed in writing before any client data is shared.

Offshore staff augmentation, slide-only strategy, penetration testing, and compliance certifications. We design, build, and hand over systems — and we leave the legal, audit, and certification decisions with your qualified counsel and auditors.