NexiaVita

The Problem With Generic AI in Clinical Work

Feed ChatGPT a patient's context and ask for a clinical note, and you'll spend more time writing the prompt and correcting the output than if you'd typed the note yourself. Generic chat puts all the contextualization work on the user — the opposite of what a busy therapist needs.

NexiaVita inverts that: the therapist opens a session for a specific patient and the system injects that patient's encrypted clinical context automatically — baseline mobility, cognitive deficits, prior goals. No prompt, no repeating themselves. They fill in the structured fields they already use, and the AI drafts from there.

AI Drafts, the Therapist Decides

In healthcare, AI can't replace professional judgment, so nothing gets written to a record silently. Every AI-generated note lands in an editable review step first. Corrections aren't errors to avoid — they're the feedback loop that makes the next draft better.

Narrowing the product to one domain, occupational therapy, made this natural: the vocabulary, the scales and the workflow match how therapists actually think, instead of asking them to adapt to a generic assistant.

Three Things That Came Out of Real Clinical Use

• Standardized scales as UI, not text. Barthel, Tinetti and similar assessments are native components — click through them and the score and interpretation are computed automatically.
• Therapy content on demand. Custom cognitive games (Sudokus, word searches) generated for the session and rendered directly in the UI.
• Grounded legal answers. A hybrid BM25 + vector search over Spanish dependency law, so regulatory questions get cited answers instead of hallucinated ones.

Where the Infrastructure Came From

Building NexiaVita surfaced a hard constraint fast: how do you send clinical notes to an LLM provider without leaking protected health information? That exact problem led to PII Firewall (detect and pseudonymize patient data before it leaves the app) and GateForge (observe cost, latency and routing without logging the sensitive payload).

What I Learned

Taking this from a chat idea to a focused vertical workflow changed how I think about AI products: the win isn't a smarter model, it's a safer, more constrained interaction — domain language, rigid boundaries, and corrections that cost the user nothing. Generic chat rarely gets you there; hiding the AI behind a workflow the user already trusts does.