Inspiration
Medical education is notoriously difficult, requiring the retention of vast amounts of complex information. Students often resort to rote memorization, which is brittle and tedious. I wanted to build a tool that doesn't just summarize notes but actively transforms them into memorable formats—like songs, rhymes, and mnemonics—mimicking the creative study hacks top students use, but automated by AI.
What it does
NeuroDose acts as an intelligent medical study companion. It takes raw medical text or documents and:
- Validates Accuracy: Cross-references user notes against trusted web sources to ensure medical facts are correct.
- Chunks Information: Intelligently segments large documents into coherent, topic-based sections.
- Adapts Learning Styles: Automatically decides the optimal memorization technique (mnemonic, rhyme, or song) based on the content's structure.
- Generates Creative Content: Produces catchy songs, rhymes, and targeted active recall questions to reinforce deep understanding.
How I built it
The core "brain" of NeuroDose is Google's Gemini 3, integrated via the google-genai Python SDK within a FastAPI backend.
- Logic & Reasoning: I utilize Gemini's multi-turn reasoning to analyze text and determine the best output format.
- Hyperparameter Tuning: I implemented dynamic temperature control ($T$) to optimize performance for different tasks:
- I use a lower temperature ($T \in [0.2, 0.3]$) for factual verification and decision-making logic.
- I use a higher temperature ($T = 0.7$) for creative generation (songs and mnemonics).
- Constraint Management: To ensure study materials are concise and memorizable, I enforce a configurable output limit of $N_{tokens} = 2048$.
- Verification Layer: I integrated Tavily to perform real-time web searches, allowing Gemini to cross-reference and hallucination-check medical data before generation.
Challenges I ran into
- Balancing Creativity vs. Accuracy: Medical data must be precise, but mnemonics need to be creative. Tuning the temperature parameters to switch between "doctor mode" and "songwriter mode" required significant testing.
- Context Management: Handling large medical documents was difficult; I had to implement intelligent text chunking strategies to keep the AI focused without losing the broader context.
- Hallucination Risks: Relying solely on LLMs for medicine is risky. Integrating the Tavily search layer was a crucial step to ensure the generated study aids were factually grounded.
Accomplishments that I'm proud of
- True Adaptive Learning: I didn't just build a wrapper; I built a system that reasons. The app autonomously analyzes content to decide if a concept is best learned via a song or a mnemonic, rather than forcing a format on the user.
- Seamless Integration: Successfully orchestrating the flow between input analysis, Tavily verification, and Gemini's creative generation in
backend/ai_utils.py. - Educational Impact: Creating a tool that transforms dry, complex medical facts into engaging, easy-to-digest content.
What I learned
I learned that prompt engineering is architecture. The quality of the output depended heavily on how I structured the multi-turn conversations with Gemini. I also gained deep insights into the importance of "temperature" variables in GenAI—learning that mathematical precision in hyperparameter tuning ($T$ values) directly correlates to the user experience.
What's next for NeuroDose
- Multimodal Support: Allowing users to upload diagrams and anatomy charts for Gemini to analyze visually.
- Audio Generation: Converting the generated lyrics and songs into actual audio files using text-to-speech or music generation models.
- Expanded Subjects: While currently optimized for medicine, the underlying logic could easily be adapted for law, engineering, or language learning.
Log in or sign up for Devpost to join the conversation.