Inspiration
We were struggling to manage our day-to-day health routines. We found ourselves constantly jumping between fragmented tools—we had dozens of separate, disconnected chats with Gemini and ChatGPT for specific advice, while our standard fitness apps just passively collected data. Nothing felt unified, proactive, or capable of actually understanding our daily lives. We realized we needed to "scratch our own itch." The inspiration for Health Twin was born out of a desire to build the perfect, all-in-one companion that tracks, remembers, and suggests actionable health habits tailored to our actual schedules.
What it is supposed to do
Health Twin is a fully autonomous, proactive AI health companion that lives right where you already chat: Telegram. Instead of waiting for you to log data, it uses a "Heartbeat Engine" to reach out to you proactively based on your personal schedule. It wakes you up with daily summaries, tracks your nutrition when you simply text it what you ate, and initiates empathetic mental health check-ins. By understanding your specific goals, biological cycles, and daily schedule, Health Twin provides highly personalized, "doable" health advice that fits your reality.
How we built it
We shifted away from traditional, rigid chatbot flowcharts and built a true Agentic architecture. We used the Google Agent Development Kit (ADK) powered by Gemini 2.5 Flash as our core Orchestrator. The ADK routes natural language requests to custom Python functions (Tools) we wrote for specific tasks. We wrapped this intelligence in a FastAPI backend, utilized Redis to manage short-term state and contextual memory, and implemented APScheduler for the background cron jobs that trigger proactive messages. Everything is containerized using Docker and deployed on an AWS EC2 Amazon Linux instance via Telegram Webhooks
Challenges we ran into
Figuring out how to properly train the agent to execute specific user tasks was a major hurdle. Moving away from rigid if/else statements to setting up the ADK orchestrator logic was completely "brain-tickling." We had to iterate heavily on our tool docstrings and system prompts so the AI would reliably know exactly when to trigger the onboarding tool versus the meal-logging tool without getting confused. Additionally, setting up the AWS infrastructure and securely exposing our local Docker containers to Telegram's strict HTTPS webhook requirements took a lot of debugging.
Accomplishments that we're proud of
While we are staying humble, getting the Telegram bot fully functional and successfully responding correctly via autonomous tool routing is a huge win for us! Watching the AI read a natural language message, decide on its own to trigger a Python script on the backend, and then return a tailored, context-aware response on Telegram was an incredible "it's alive!" moment for our team.
What we learned
We learned how to think in "Agentic workflows"—learning to trust the LLM to act as a reasoning engine rather than just a text generator. We also leveled up our backend engineering skills, figuring out how to maintain stateful, context-aware conversations across stateless HTTP webhook requests using Redis, and getting comfortable with AWS EC2 server management and Docker deployments.
What's next for Health Twin Agent
This MVP is just the foundation. Our immediate next steps include:
- Snowflake Integration: Replacing our Redis profile cache with a robust Snowflake data warehouse for long-term health profiling and trend analysis.
- Multi-Agent Architecture: Using the ADK to build specialized Sub-Agents (e.g., a "Lab Results Analyzer" or an "IBS Tracker") that our Root Agent can delegate complex queries to.
- Voice Interactivity: Integrating voice-to-text so users can do hands-free journaling and logging while on the go.
Log in or sign up for Devpost to join the conversation.