What we do

Building and evolving production web applications: architecture, stack selection, delivery, and the refactoring that keeps a codebase alive years later. Front-end at the core, plus the back-end a product needs around it, the APIs, data models and integrations behind the surfaces we ship. This is the spine of everything else on this page.

Products where the model is the product: vision-language pipelines, LLM analyst workflows, avatar and dubbing generation, and platforms where your own people author agents. Four independent engagements, different clients, different model classes, not a pivot.

Setting your own people up to work with AI rather than just shipping you AI features: pipelines, tooling, working practices, and teaching efficient use. That includes leaving a repository your non-technical people can go on shipping from, in-repo skills, hooks and agents that turn "add this section" or "adjust this view" into something you do yourself.

We design the surfaces we build, working AI-first: Claude-driven design workflows that go from a problem statement to a working, styled interface without a separate design agency in the loop.

We place a technical lead who owns architecture decisions, sprint planning, developer coordination and the client conversation, working alongside your engineers rather than behind a vendor wall. Five of our people can hold that role, so it is a capability rather than one person's calendar.

Product surfaces where the difficulty sits behind the UI: financial state that has to be exactly right, compliance gates, live on-chain data, medical data under HIPAA. And the dashboards professionals sit in all day, live mission control, surveillance review, portfolio and risk analytics, configurable layouts each user assembles for themselves.

Taking over a large application somebody else built badly, working out by hand what is load-bearing, and returning it to a state where a redesign and new features are possible again. The claim is the whole sequence, not just the refactor: map, plan, rewrite, upgrade, then actually ship features on the result.

Authoring platforms where non-engineers build lessons and tests and the result behaves like software: an activity model rich enough for puzzles, gap-fills and drag-and-drop, role separation between author and learner, and completion that can be proven afterwards.