OncoWispr

Inspiration

We heard countless stories of cancer patients who did not have the physical or mental energy to type out their feelings, track symptoms, or find resources while adjusting to aggressive therapy. We built OncoWispr to mitigate this exact issue of turning stream-of-consciousness spoken journal entries into actionable clinical insights and low-energy updates for loved ones.

What it does

OncoWispr is an intelligent, real-time clinical dashboard ecosystem powered by advanced NLP processing. When a patient speaks into the microphone, they see these exact features play out in front of their eyes:

1. The Live Stream Monitor: Instantly captures and parses the raw text stream.
2. The Predictive Analytics Engine: Evaluates conversational speech for emotional distress densities and extracts different signals from their speech, such as tiredness or anxiety.
3. The Smart Caregiver Summary: Uses LLMs to strip out complex medical terminology and automatically generate gentle, bite-sized status updates for family members and caregivers.
4. The Adaptive Resource Routing: Automatically changes state based on live health metrics. If an elevated stress vector is detected, the dashboard's design token system dynamically morphs, shifting buttons to high-priority red states, updating critical banners, and rendering urgent access lines like the 988 Lifeline.

How we built it

Our system utilized a dual-layered framework featuring a: Frontend: Built with distinct tools such as React and Vite, styled with a high-contrast dark-mode glassmorphic theme. We mapped a distinct brand palette (#D84231 for urgent alerts, #D1FAD5 for backgrounds, #4FB58C for lines/highlights, and #1AD82C for normal baseline statuses) to provide immediate visual feedback. Database & Stream: Backed by Cloud Firestore. The frontend maintains live snapshot listeners via secure WebSockets to catch cloud changes instantly without requiring a page reload. Intelligence Layer & Backend: Powered by the Groq SDK utilizing the Llama 3 (70B) model. Our backend pipeline structures raw transcription data into strictly typed JSON objects containing sentiment tags, low-energy summaries, and parsed biomarkers. Authenticated write tokens securely pass this directly into the cloud.

Challenges we ran into

One major hurdle was establishing a completely secure, asynchronous data bridge between backend recording modules and our web interface. We initially drafted a standalone Node setup (pipeline.js), but safely migrating the administrative Master Handshake to a backend script using a private Service Account JSON configuration allowed us to completely prevent secret key leaks while enabling direct cloud injection. Additionally, fine-tuning the prompt engineering parameters to guarantee strictly valid JSON objects from the Llama 3 context required careful formatting guardrails.

Accomplishments that we're proud of

We are incredibly proud of our Dynamic Interface Linking. Seeing the entire theme, badge system, and button configurations of the Community Resources module organically morph from a calm green baseline state into an urgent, actionable red state purely because a patient voiced physical exhaustion feels like a massive step forward for empathetic UX design.

What we learned

We learned the importance of true full-stack coordination and why server-side execution layers belong completely separated from client-side environments. Building this project also gave us deep insight into how NLP can bridge the communication gap between intensive clinical treatments and caregiving families.

What's next for OncoWispr

We plan to scale our predictive health model into a dedicated mobile ecosystem using React Native. We want to implement long-term vector embeddings via tools like Pinecone to track subtle semantic drift across multiple months of speech, allowing healthcare providers to catch gradual depressive or physiological warning signs much earlier in the recovery process.

Built With

• claude
• firebase
• groq
• html
• javascript
• node.js
• python
• react-native
• supabase