Reminiscence AI

Reminisce AI – The Cognitive Anchor A personalized voice-based AI companion designed to provide comfort and cognitive support to individuals with dementia by leveraging their unique life stories.

📖 Inspiration The inspiration for Reminisce AI stems from the profound challenges faced by families caring for loved ones with dementia. We noticed that while medical care is essential, the emotional and cognitive "anchoring" provided by familiar stories and voices is often missing in daily care. We wanted to build a "Cognitive Anchor"—a companion that doesn't just talk, but remembers a patient's specific life history to provide comfort during moments of confusion or loneliness.

✨ What it does Reminisce AI acts as a voice-driven companion that bridges the gap between memory and interaction:

Onboarding: Caregivers provide a detailed biography and choose a soothing, familiar voice for the AI.

Contextual Conversations: The AI uses this biography to engage in meaningful dialogue, referencing past experiences, family members, and favorite memories.

Emotional Grounding: By using familiar context, it helps ground the patient in reality and significantly reduces anxiety.

🛠️ How we built it We engineered a robust AI stack focused on low-latency response times and high emotional intelligence:

Frontend: Developed with React to create a simple, high-contrast, and accessible interface for elderly users.

Speech-to-Text (STT): Integrated OpenAI's Whisper (base model) for accurate real-time transcription of patient speech.

The Brain (LLM): Leveraged Google's Gemini 2.5 Flash (and 2.0 Flash) to process life stories and generate empathetic, context-aware responses.

Text-to-Speech (TTS): Utilized ElevenLabs API to synthesize responses into warm, human-like voices.

Backend: A FastAPI server orchestrates AI services and maintains persistent memory via a local JSON database.

🚧 Challenges we faced Latency Sync: Ensuring the transition from patient speech to AI response felt natural was a major hurdle. We optimized the backend to handle processing tasks in parallel.

API Rate Limits: We encountered 429 Resource Exhausted errors during heavy testing. We solved this by implementing a robust fallback system that provides immediate comforting messages if the API is throttled, ensuring the patient is never met with silence.

Context Persistence: We built a custom JSON persistence layer to ensure the AI "remembers" the patient setup even after server restarts.

💡 What we learned We learned that in healthcare AI, "personality" is just as important as "accuracy". Tuning the model's tone to be patient, loving, and repetitive (when necessary) was vital in making the tool feel like a true companion rather than a cold computer interface.

🚀 Tech Stack Frontend: React

Backend: FastAPI (Python)

AI Models: Gemini 2.5 Flash, Whisper STT

Voice: ElevenLabs TTS

Storage: Local JSON Persistence

Built With

• fastapi-(backend)-ai-models:-google-gemini-2.5-flash
• javascript-frameworks:-react-(frontend)
• languages:-python
• openai-whisper-apis:-elevenlabs-(voice-synthesis)-tools:-dotenv-(environment-management)
• uvicorn