GreenWatch; AI-Powered Fatigue Risk Detection Intelligence

Inspiration

GreenWatch was inspired by months of research into fatigue, stress, and mental health challenges in safety-critical industries - particularly maritime environments. Through our collaboration with an international research partner, we uncovered how often cognitive fatigue and emotional strain go unnoticed until they lead to serious accidents, financial loss, or personal tragedy. We wanted to build a system that could proactively detect early warning signs and intervene before issues escalate.

What it does

GreenWatch is a local-first AI fatigue-risk detection system designed for safety-critical workplaces.

It: Engages workers in natural, low-pressure conversations Analyzes responses using a small on-device model Detects early signals of stress, fatigue, or cognitive overload Logs structured risk data Stores embeddings in Actian VectorAI DB for semantic search and trend analysis Flags elevated risk levels for proactive intervention

The goal isn’t just emergency response - it’s early mitigation to reduce burnout, prevent accidents, and improve long-term operational safety.

How I built it

We built GreenWatch as a full-stack system: Frontend: React + Tailwind browser app for conversational UI Inference Layer: Local small model running via ExecuTorch Runner: C++ inference pipeline handling model execution and scoring Database: Actian VectorAI DB for structured logs and vector embeddings Containerization: Dockerized environment for reproducibility The system runs locally, ensuring privacy and low-latency inference while maintaining structured risk tracking and semantic memory through VectorAI.

Challenges I ran into

Integrating ExecuTorch with a C++ runner under hackathon time pressure Learning and implementing Actian VectorAI DB from scratch Connecting frontend → local inference → structured logging → vector storage cleanly Managing GitHub merge issues while building across multiple components Designing conversations that feel supportive instead of overwhelming Balancing technical complexity with usability was one of the hardest parts.

Accomplishments that I'm proud of

Shipping a fully integrated AI system in 48 hours Successfully implementing Actian VectorAI DB with semantic embeddings Running local on-device inference with ExecuTorch Winning 1st Place – Best Use of Actian VectorAI DB Building something grounded in real research with real-world impact Most importantly, we moved beyond a prototype chatbot and created a structured, scalable risk-detection framework.

What I learned

How to integrate vector databases into AI safety workflows How on-device inference changes system architecture decisions The importance of modular system design under time pressure That building meaningful AI requires balancing engineering, ethics, and UX How powerful rapid experimentation can be when paired with strong teamwork

What's next for GreenWatch

Next steps include: Improving model accuracy with better fine-tuning for fatigue detection Expanding risk scoring into longitudinal trend analysis Adding multimodal inputs (e.g., biometrics, shift data) Running pilot programs with industry partners Exploring enterprise deployment and investment opportunities Long-term, GreenWatch aims to reduce preventable accidents, mitigate burnout early, and save organizations billions in fatigue-related operational losses - while protecting the people who keep critical industries running.

Built With

• ai
• c++
• executorch
• javascript
• native
• react
• vector