Projects

Aura: Closed-Loop AIoT Hardware & Inference System

Edge-to-cloud pipeline bridging cloud LLMs with local IoT hardware for real-time, sentiment-driven physical actuation: inline control tags are parsed mid-stream from raw SSE responses to drive motor commands as the model streams. Reverse-engineered the proprietary BLE protocol by hand-parsing raw Bluetooth HCI captures (btsnoop → ACL → L2CAP → ATT framing) without any BLE libraries, using statistical differential analysis and checksum hypothesis testing to map unknown encrypted hex payloads to precise motor frequencies. Fronted by a cost-aware LLM gateway on Cloudflare Workers with zero-buffer SSE passthrough, cutting gateway-added latency by ~92% (399ms → 30ms). Extended with a haptic meditation module delivering programmable breathing patterns (e.g., 4-7-8 cycle) via vibration motor, synchronized with streaming audio cues. Currently in closed beta.

  • Python
  • HTML/CSS
  • Cloudflare Workers
  • Bluetooth HCI · L2CAP · ATT
  • SSE Streaming
  • LLM APIs
  • Edge Inference
  • Haptic Feedback

Hybrid AI Safety Agent for Fall Detection

Won highest score in Solos Technology internal hackathon. Autonomous 4-layer fall detection system for smart glasses: IMU impact detection triggers VLM-based visual confirmation (first-person camera), followed by a 30-second recovery cancel window. On timeout, an automated alert with GPS location is sent to emergency contacts. BLE disconnection during a fall event is treated as severity confirmation. Built a lock-free ring buffer for continuous pre-fall frame capture, enabling reliable photo evidence even when the device is face-down after impact.

  • Kotlin
  • Gemini VLM
  • n8n
  • BLE
  • Android
  • Edge AI
  • Solos V2 SDK

Local Multimodal Semantic Video Search

Zero-API local RAG pipeline using OpenCV for frame extraction and LLaVA for semantic captioning, with cross-lingual query routing via Gemma 2 and Nomic embeddings. Profiled Python GIL bottlenecks (10:1 compute ratio) and prototyped a native C++ ingestion layer via OpenVINO, achieving 2.8× throughput at 4× sampling density over the Python baseline.

  • Python
  • C++
  • OpenVINO
  • OpenCV
  • LLaVA
  • Ollama
GitHub →

Kaggle: Drawing with LLMs

Bronze Medal — ranked 103rd / 1,309 globally (top 8%). Designed prompt-engineering pipelines that guide frozen LLMs to generate precise, executable Python for complex image reconstruction tasks.

  • Python
  • Prompt Engineering
  • LLMs