Inspiration
I'm Oluwatominiyi, a solo developer and engineering student in Nigeria. 55 million people have dementia worldwide. Most are diagnosed 2-3 years too late. In underserved regions, the tools to catch it early are expensive or don't exist. Nobody goes to a hospital because they forgot their keys, but they might talk to an app for 30 seconds. That's CogniCare.
What it does
CogniCare detects early signs of cognitive decline through daily voice conversations and memory games. Users record a 30-second voice note answering a simple question. The AI analyzes speech patterns (pauses, filler words, pitch, speech rate) and combines it with a memory game score to produce a Brain Health Score (0-100) with color-coded alerts. A family dashboard lets caregivers monitor loved ones and get notified of significant changes.
How we built it
Backend (Python FastAPI): Whisper for speech-to-text, Librosa for acoustic feature extraction (pitch, pauses, speech rate), rule-based scoring derived from clinical research, FFmpeg for audio format conversion.
Frontend (React + Tailwind): Voice recorder with waveform, circular score gauge, trend line chart, 4-card memory game, family dashboard with alert panel. Senior-friendly design: 18px+ fonts, high contrast, large touch targets, warm colors.
Challenges we ran into
- Audio format mismatch: Browsers record webm, but Whisper needs WAV. Fixed with FFmpeg conversion on the server.
- CPU-only inference: Whisper runs slow without GPU. Acceptable for demo, needs optimization for production.
- No training data: Built rule-based scoring from research papers instead of ML. Ready for real data when available.
- False alerts: Early versions flagged healthy speech. Tuned thresholds manually through repeated testing.
- Solo build: Every decision made alone. Prioritized what mattered most for the demo.
- Stigma-aware design: Naming, colors, and language all affect whether people actually use it.
Accomplishments that we're proud of
- Built a complete working product, not a slide deck
- End-to-end AI pipeline: voice recording to cognitive score in under 30 seconds
- Senior-friendly UI that judges could imagine their parents using
- Family alert system that shows the real-world value of early detection
- Finished as a solo builder with a working demo
What we learned
Speech biomarker extraction is technically challenging but the real problem is human: people avoid cognitive tests because of stigma. A warm, simple, privacy-first design matters more than algorithmic accuracy. Also: audio format handling, CPU ML inference tradeoffs, and rule-based clinical scoring when training data doesn't exist.
What's next for CogniCare
This is a minimal working version. Next steps: validate scoring against clinical data, add multi-language support, improve Whisper inference speed, build HIPAA-compliant storage, and partner with healthcare providers for real-world testing. Winning this hackathon helps make that happen.
Log in or sign up for Devpost to join the conversation.