Inspiration
Not everyone experiences stress the same way. Individuals who have experienced trauma or live with conditions such as anxiety and PTSD may have unique triggers and physiological responses that aren't always apparent to others. We wanted to create a private, judgment-free space where people could express themselves while gaining a better understanding of their own physiological responses.
What it does
CollAI is an offline AI venting companion that allows users to have private conversations without relying on cloud-based AI services.
While interacting with CollAI, users can monitor their heart rate in real time and visualize physiological trends alongside their conversations. CollAI also maintains conversation histories, allowing users to revisit previous interactions and observe potential patterns between conversational topics and physiological responses.
Users can optionally share these records with healthcare professionals to support discussions about their well-being. CollAI is not intended to diagnose conditions or replace professional healthcare.
How we built it
We built CollAI around the Arduino UNO Q, deploying a locally running generative AI language model through Ollama. We used llama.cpp and GGUF quantization to prepare our model for local deployment.
For biometric monitoring, we integrated a MAX30105 heart rate sensor with the UNO Q. We used MATLAB to develop real-time heart rate monitoring and signal visualization, allowing users to observe changes in their physiological readings. Python was used for conversation management, data logging, lora training, and dashboard development. We also implemented JSONL storage to maintain conversation histories and their associated heart rate readings.
Challenges we ran into
One of our biggest challenges was integrating several different systems into one functioning prototype. We encountered difficulties with sensor reliability, noisy heart rate measurements, local model deployment, audio integration, and communication between the Arduino UNO Q, MATLAB, and Python applications.
We also had to work through the limitations of running a quantized language model on embedded hardware while maintaining a responsive conversational experience.
Accomplishments that we're proud of
We're proud to have developed a working prototype that brings together offline generative AI, biometric sensing, and physiological data visualization. We successfully deployed a locally running language model, integrated heart rate monitoring, implemented conversation history logging, and developed MATLAB and Python tools to visualize physiological trends.
More importantly, we're proud of the problem we're addressing. Privacy and individual differences in stress responses are often overlooked in conversations about mental well-being.
What we learned
Throughout this project, we gained hands-on experience with local LLM deployment, model quantization, embedded Linux, biometric sensor integration, Python development, MATLAB-based signal visualization, and hardware-software communication.
We also learned that developing technology for mental well-being requires more than building something technically functional. Privacy, accessibility, responsible data interpretation, and user autonomy must be considered throughout the development process.
What's next for collai
We want to improve heart rate measurement reliability, refine the conversational experience, and expand our physiological data visualization capabilities.
Future development will focus on making conversation histories and heart rate trends easier to interpret and export, allowing users to share information with healthcare professionals if they choose.
Ultimately, we hope to develop CollAI into a customizable, privacy-first companion that empowers people to better understand their own experiences without compromising their personal data.
