AI systems
CabinAI
A seven-agent driving-assistant prototype combining local perception, voice, and retrieval with cloud reasoning. Jury Award winner at Qualcomm SparQ 2026.
Play demo CABIN AIEDGE ↔ CLOUD
7 agents. One shared intelligence. SparQ winner
About the project
CabinAI combines driver-state detection, location, voice, and retrieval through seven agents. Local perception and safety rules feed a shared orchestration layer. A router handles supported requests on-device and escalates complex questions to cloud models. I built and demonstrated it solo at Qualcomm SparQ 2026. Fatigue forecasting remains experimental.
What I built
- Architected a 7-agent hybrid edge-cloud vision-language-action system: safety and voice run on-device, while complex reasoning can escalate to Qwen3-VL-32B / gpt-oss-20b on Qualcomm AIC100.
- Designed a core drowsiness detection pipeline utilizing a 3-engine vote consensus (rule-based, ML, and LLM) to drastically reduce false positives.
- Developed ZeroClaw, a state orchestration bus tying all agents with a location-aware voice assistant into a shared, proactive intelligence layer.
- Demonstrated the system end-to-end via a full driving simulation through Gachibowli, Hyderabad.
- Jury Award Winner at SparQ Qualcomm TechFest 2026 , competing solo against 166 teams.
- Implemented a deterministic edge/cloud router with multi-tier fallbacks. Routing 80%+ of queries locally was the design target; the distribution depends on the workload.
- Implemented local retrieval over a 31-document corpus with BGE-small ONNX embeddings and a keyword fallback, without a PyTorch runtime dependency.