Hexi

Inspiration

Clinical documentation is one of the largest hidden burdens in healthcare. Nurses often spend hours typing notes and may record notes that lack key context, contributing to the growing problem of unstructured data in healthcare. Concurrently, most AI scribes and copilots send raw patient conversations to cloud models — creating serious privacy, compliance, and data ownership concerns.

We asked a simple question: What if we could build a clinical copilot that improves documentation quality while making privacy the default, not a feature?

Hexi solves this: it reduces the burden of documentation, identifies missing clinical information in real time, and ensures that sensitive patient data never leaves the browser unprotected.

What it does

Hexi is a real-time, privacy-first clinical conversation copilot.

It:

• Transcribes patient conversations fully in the browser.
• Runs a custom 350M parameter LoRA redaction model locally in the browser to detect and censor personal health information.
• Sends only redacted text to state-of-the-art cloud reasoning models, such as Claude Sonnet.
• Identifies missing clinical information (HPI gaps, medications, allergies, history, etc.).
• Guides structured follow-up questions, leading to the construction of a complete patient profile.
• Knows when to stop asking questions once enough information has been gathered.

Unlike typical AI scribes that simply summarize transcripts, Hexi actively detects knowledge gaps and helps clinicians complete structured documentation efficiently. Most importantly, no personal health information is ever sent to the cloud.

How we built it

Our architecture is privacy-first by design:

• For our frontend, we use React + TypeScript for type safety.
• In-browser model execution using ONNX/WebGPU.
• A custom LoRA fine-tuned 350M parameter model trained using transfer learning from Claude Sonnet 4.
• Real-time streaming transcription that happens completely on device.
• Redacted text is sent to a state-of-the-art cloud reasoning engine (Claude Sonnet configured as a multi-turn reasoning agent).

The redaction model we used is a state-of-the-art "Liquid Foundation Model" by Liquid AI. These models are small yet offer excellent performance and work well in edge deployments.

We also implemented a multi-turn agent framework that:

• Tracks clinical slot completion.
• Measures marginal information gain.
• Prevents infinite follow-up loops.

This allows Hexi to stop when sufficient information has been gathered instead of asking endless general questions.

Challenges we ran into

• Running a 350M parameter model efficiently inside the browser.
• Combining regex-based redaction with model-based entity detection.
• Preventing the agent from asking overly general or repetitive questions.
• Implementing a formal stopping condition for “enough information.”
• Handling malformed JSON outputs from the reasoning model.
• Maintaining a smooth real-time UX while models load locally.

Accomplishments that we're proud of

• Successfully running a 350M parameter redaction model entirely in-browser.
• Training an accurate model using transfer learning from Claude Sonnet, resulting in a scalable approach.
• Achieving full privacy-by-default architecture (only redacted text leaves the device).
• Building a working multi-turn clinical reasoning agent.
• Implementing a measurable information sufficiency stopping mechanism.

In a short timeframe, we built not just a demo, but a functional system with real architectural integrity.

What we learned

• Edge AI in the browser is more viable than most people think, but it requires careful performance tuning.
• Multi-turn agents need explicit structure; without constraints, they drift into generic questioning.
• Measuring “enough information” is just as important as generating answers.
• Privacy-first design simplifies downstream trust and compliance conversations.
• Clear UX is absolutely essential, especially in rushed healthcare environments.

What's next for Hexi

• Clarify how we would integrate with patient-led medical record storage systems such as Healthnix.
• Explore integrations with EHR systems to prevent the double-entry problem, where nurses first enter information in our software and then in the EHR.
• Refine the information sufficiency scoring system to prevent redundant follow-ups and better characterize what is "enough."
• Optimize model loading speed on mobile browsers.
• Conduct clinical usability testing with nurses to assess the real-world impact of our product.

Our long-term vision is to make Hexi the default clinical copilot for nurses — one that enhances care quality while protecting patient privacy at its core.

Built With

• claude
• liquid-ai
• lora
• machine-learning
• multi-turn-reasoning
• python
• react
• transfer-learning