Vitena is a practice-management platform for health and wellness professionals, paired with a public directory where clients find and book them. It runs in production with real users, in 7 languages, on the web and as an Android app.
- Provider-agnostic LLM layer
- One completion interface in .NET with an adapter per provider. Model and cost policy are set per task type; every call is metered per tenant and written to a usage ledger with its real cost.
- Agents in production
- A tool-calling agent loop behind three chat surfaces: an in-app support agent with read-only account tools, a public pre-sales assistant, and an admin agent that answers questions through read-only SQL with a statement timeout and a row cap.
- Retrieval without embeddings
- A document map in the system prompt plus a
read_doc tool over the product's AI-facing documents, with the prompt ordered so that most input tokens are served from cache.
- 20+ LLM task types
- Meal-plan generation where nutrition values are computed in code, not by the model. Session-note drafts, document reading with vision, voice-to-record capture, triage of email and feedback, moderation checks.
- Tracing and evals
- Tracing at the adapter layer (prompt, response, tokens, cost, latency; content withheld for health data), a spend report by task type, and a black-box eval harness that runs whenever a prompt changes.
- Public MCP server
- A read-only MCP server over Streamable HTTP with 3 tools, listed in the official MCP Registry as
care.vitena/directory.
- Readable to AI assistants
- Server-rendered and prerendered pages, JSON-LD, a sitemap and
llms.txt, with server-side analytics for AI bots.
- Coding-agent pipeline
- A multi-session pipeline on Claude Code: a server-side state machine advances the subtasks of a work item and a local bridge opens one headless session per step, with one upfront plan approval and permission-gated deploys.