MediSummary
Transforming messy clinical notes into professional documentation and patient clarity with Gemini 3.0 Flash.
About the Project
Inspiration: The "Invisible Patient"
In modern healthcare, there is a silent crisis known as the "2:1 ratio." For every hour a physician spends face-to-face with a patient, they spend nearly two hours tethered to a keyboard, documenting the encounter. This administrative burden is a primary driver of physician burnout and, more importantly, it makes the patient feel "invisible" during their own appointment.
We built MediSummary to turn the doctor's focus back to the human in the room. By leveraging the low-latency reasoning of Gemini 3.0 Flash, we aimed to create a "Clinical Intelligence Partner" that listens to the raw, messy narrative of a consultation and instantly structures it into professional, actionable data.
How We Built It
The project is architected as a full-stack workspace designed for speed and clarity:
- The Brain: We utilized the
gemini-3-flash-previewmodel via the Google Generative AI SDK. Its ability to process long-context clinical notes while maintaining high reasoning capabilities allowed us to implement complex features like SOAP Note generation and Checklist extraction with minimal prompt tuning. - The Interface: We moved away from traditional "clinical" (cold/white) UIs. Instead, we developed a "Cosmic Glass" aesthetic using React 19 and custom CSS. This dark-mode, bento-grid layout reduces eye strain for providers working night shifts and prioritizes information hierarchy.
- The Accessibility: We integrated the Web Speech API to provide Text-to-Speech (TTS), allowing doctors to listen to summaries while prepping for their next patient, effectively multitasking without increasing cognitive load.
Mathematical Efficiency
The core value proposition can be viewed through the lens of time-to-documentation ($T_d$): If $T_m$ is the time spent on manual documentation and $T_a$ is the time spent on AI-assisted documentation, we observe: $$T_a \approx 0.15 \times T_m$$ By reducing the documentation overhead, we increase the Patient Interaction Ratio ($R_p$): $$R_p = \frac{T_{patient}}{T_{doc} + T_{patient}}$$ Our goal was to move $R_p$ from $0.33$ toward $0.85$.
Challenges We Faced
- The "Jargon Gap": One of our biggest hurdles was "Patient Mode." We had to engineer prompts that could translate high-level medical terms (e.g., "Myocardial Infarction") into empathetic, non-alarming language ("A situation where your heart muscle needed more oxygen") without losing the medical urgency.
- Structuring the Unstructured: Clinical notes are often a "stream of consciousness." Ensuring the AI correctly categorized data into the SOAP (Subjective, Objective, Assessment, Plan) format required rigorous system instructions to avoid "hallucinating" vitals that weren't present.
What We Learned
We learned that AI in healthcare isn't about replacing the doctor's judgment; it's about removing the friction of the tools they use. We discovered that small UX details—like a "Read Aloud" button or a "One-Click Copy"—can be just as impactful as the AI model itself when it comes to real-world clinical adoption.
Log in or sign up for Devpost to join the conversation.