Devpost Link: https://devpost.com/software/dear-me-8emch1

🛠️ Tech Stack

🗣️ XTTS-V2, Ollama Mistral (local) 📂 Redis, Chroma
🎨 Next.js, React, TailwindCSS
⚙️ Python, FastAPI
📝 Custom algorithms
📴 Local device support

Inspiration

The inspiration for Dear Me came from the idea of making self-reflection more accessible and meaningful in today’s busy world. Journaling is a proven way to boost mental well-being and track personal growth, but many people struggle to maintain the habit... (like us who built DearMe)! We wanted to create a solution that feels natural, effortless, and secure, encouraging people to reflect without added stress.

What it does

Dear Me is a private, offline conversational AI that acts as your personal companion throughout the day. By running on a local LLM, speech models, and database, it ensures all your data belongs to yourself, for yourself! You can share your thoughts, feelings, and experiences in natural spoken and written conversations throughout the day, and at the end of the day, it generates a personalized journal entry summarizing your interactions.

How we built it

We designed and built Dear Me with a strong focus on privacy and usability:

• AI Model: Two locally fine-tuned LLMs for natural conversations and journal entries. Both LLMs use Retrieval Augmented Generation and accurate prompt engineering.
• Voice Text-to-speech: Enabled real-time speech-to-speech conversations by leveraging XTTS-V2 locally.
• Database: A dockerized local Redis database to keep information private. Also we used a local Chroma Database, made with embeddings of hand-picked research papers, articles, and blogs for Retrieval Augmented Generation.
• Frontend: Built with NextJS/React and TailwindCSS to deliver a clean and intuitive user experience.
• Backend: Developed using Python and FastAPI for efficient conversational flow and local data handling.
• Summarization Logic: Designed algorithms to extract meaningful insights and craft journal entries in a cohesive, personalized style.
• Offline Optimization: Focused on ensuring the app runs smoothly on local devices without compromising performance.

Challenges we ran into

• Fine-Tuning the Model: Balancing conversational naturalness with effective summarization required extensive experimentation.
• Local Performance: Ensuring the AI runs efficiently on diverse hardware setups while staying offline was technically demanding, required a lot of testing.
• User Experience: Designing an interface and conversational flow that felt engaging yet simple took multiple iterations.

Accomplishments that we're proud of

• Successfully creating an offline, secure conversational AI that protects user privacy.
• Building an engine capable of transforming conversations into meaningful, personalized journal entries.
• Designing an intuitive and polished user interface that makes the app enjoyable to use.

What we learned

• How to fine-tune and optimize LLMs for specific tasks like summarization and conversational flow.
• First time using FastAPI + React tech stack, but learned swiftly to make & test API endpoints.
• Front-end standard design practices to make the app scalable.
• GIT GIT GIT

What's next for Dear Me

We have exciting plans for Dear Me in the future:

• Customization Options: Let users personalize the tone and style of their journal entries.
• Insights Dashboard: Provide users with emotional trends and reflections over time, like a "monthly wrapped".
• Voice Integration: Complete hands-free, voice-based interaction for easier access.
• Mental Health Features: Introduce mindfulness tips or guided prompts based on user interactions.

With these features, Dear Me will continue to empower users to reflect, grow, and cherish their personal stories—all while keeping their data private and secure.