Inspiration
I’m Ira Divinsky, founder of IntelliGene AI. After about 20 years in federal healthcare IT, including work with HHS, I’ve seen how difficult it can be for patients and caregivers to understand hospital discharge instructions.
Terms like “PO BID,” medication restrictions, and follow-up requirements arrive at an already stressful moment. I wanted to explore a tool that could turn those papers into clearer steps—and help hospital leaders understand which instructions people still find confusing.
What it does
Discharge Instructions Explainer demonstrates how discharge information can become:
- A plain-language summary and numbered home-care steps
- A medication schedule with doses and timing
- Separate emergency and call-your-doctor warning sections
- Questions for a follow-up visit
- Explanations in English, Spanish, Mandarin, Arabic, and Russian, including right-to-left Arabic layout
I also built a feedback loop. Users can rate the clarity of a synthetic example and flag confusing sections. An authenticated leadership dashboard aggregates those ratings, and NVIDIA Nemotron generates a short briefing with proposed interface improvements.
The current prototype accepts only two built-in fictional discharge cases. Their multilingual explanations are prepared, editorially corrected examples, clearly labeled as demonstrations rather than live inference. Custom patient documents and free-text feedback are disabled.
How I built it
I used NVIDIA Nemotron 3 Super (nvidia/nemotron-3-super-120b-a12b) through Nebius Token Factory’s OpenAI-compatible API. During development, I used it for simplification, structured output, translation, and source-checking experiments. The current leadership briefing feature uses live inference on aggregate feedback counts.
Nebius handled model serving, so I did not provision GPUs, deploy model weights, or manage inference infrastructure myself.
A Cloudflare Worker serves the HTML/JavaScript interface and API routes. The Nebius API key stays in a server-side Worker secret. Cloudflare Workers KV stores structured feedback with a 30-day expiration.
For the private leadership workflow, I added Cloudflare Access with an owner-only MFA policy and independent signed-token verification inside the Worker. I also added request-size limits, strict input validation, security headers, and protections against caching sensitive responses.
I used Grokbot and OpenAI Codex as development assistants for implementation, debugging, testing, and review.
Challenges I faced
My biggest challenge was preserving meaning while simplifying medical language. Testing showed that the model could add unsupported advice, omit warning signs, or produce inconsistent structures across languages.
I experimented with stricter prompts, JSON-mode output, runtime validation, and a second model pass to check explanations against the source. That checker sometimes withheld explanations and sometimes missed omissions. I learned that a model checking another model response is not a guarantee of correctness.
Latency was another challenge. Generating multilingual output and running additional checks made the experience less predictable. I introduced prepared synthetic examples so the demonstration could show language switching reliably without presenting that speed as live inference performance.
Privacy also required changes. I removed free-text feedback, restricted leadership access, and blocked arbitrary discharge text. These safeguards reduce exposure, but they do not establish HIPAA compliance.
Accomplishments I’m proud of
I built a working prototype that connects a patient-facing explanation interface with a feedback-to-leadership workflow.
I’m particularly proud of the multilingual presentation, separate warning categories, and the ability to turn aggregate clarity ratings into a leadership briefing using NVIDIA Nemotron on Nebius.
I also built automated tests for all ten sample/language combinations, input restrictions, feedback retention, authentication, private routes, and browser security controls.
What I learned
I learned how quickly managed model inference can support an application—and how much work remains between a convincing demonstration and a dependable healthcare product.
Structured JSON helps with presentation, but it does not establish factual accuracy. Translation needs its own evaluation. Privacy requires enforceable data boundaries, not just a disclaimer. Clinical correctness, security, and compliance each require separate evidence.
What’s next
Before considering real patient use, I would pursue:
- Clinician review of medication and warning-sign fidelity
- Professional translation and readability evaluation
- Deterministic checks for doses, thresholds, and restrictions
- Vendor agreements and a documented security/privacy assessment
- A separately reviewed patient-storage and hospital-access design
- Read-aloud support and printable instructions
This is a synthetic-data prototype, not medical advice, and I do not represent it as HIPAA compliant.
Log in or sign up for Devpost to join the conversation.