Zhongquan
Cheng.

• —Designed a distributed request pipeline on Server-Sent Events + Vercel serverless functions, supporting 50+ concurrent users with zero blocking on long-running tool calls.
• —Built an agentic chatbot with LLM tool-calling that orchestrates multi-step workflows — database lookups, form filling, and third-party API calls — end-to-end.
• —Shipped a fault-tolerant LLM gateway with latency-aware A/B provider selection and automated failover, holding 99.9% uptime while optimizing token spend during outages.
• —Cut 40% of redundant Postgres queries with a composite-key cache; enforced defense-in-depth via Row-Level Security and API middleware scoping Pinecone searches to counselor_id for zero cross-tenant leakage.

Node.js / Supabase / PostgreSQL / Vercel / Pinecone / Redis / LangChain / LLM Tool-Calling

• —Built a medical-imaging privacy pipeline on Meta Sapiens + OpenCV — 16-class semantic body-part segmentation across 10,000+ frames — with dynamic exposure adjustment and temporal quality analysis, raising low-light detection rate by 80%.
• —Designed a LangGraph multi-agent sequence-error-detection system (Perception / Validator / State-Tracker / Supervisor) with a fast-path (vLLM 4B) + slow-path (vLLM 8B) dual loop for real-time surgical-action-compliance monitoring.
• —Served Qwen3-VL in a dual-model cascade via vLLM with FP8 quantization, PagedAttention, and prefix caching — a 16-thread parallel pipeline cut per-clip latency from 200 ms → 120 ms.
• —Adapted a Video-STaR-inspired staged training recipe (SFT → hard-negative mining → temporal extension → tool augmentation) on Qwen 27B + LoRA, lifting HAR accuracy from ~75% to 85.47% on 79,600 balanced samples.
• —Proposed DeAR — Decomposing Attention Head Roles — classifying ViT heads by entropy and injecting learnable attribute tokens + role-attention masks to strengthen weak-class recognition.
• —Added a YOLO ROI crop stage upstream of the VLM, removing ~70% of background noise and sharpening hand–object interaction recognition.
• —Accelerated video ingestion with FFmpeg-based high-speed frame extraction, cutting preprocessing latency by 60% across 50+ hours of clinical footage.
• —Shipped a FastAPI real-time dashboard with SSE streaming, a HITL review queue, and ThreadPoolExecutor concurrency — handling 12,000+-frame long videos end-to-end.

Python / PyTorch / vLLM / Qwen3-VL (4B / 8B / 27B) / LangGraph / LoRA / YOLO / Meta Sapiens / OpenCV / FFmpeg / FastAPI / Docker

• —Built a scalable Android data-collection framework using the Factory design pattern to decouple sensor implementations, so IMU + audio streams can be added without touching the collection core.
• —Containerized the ML training pipeline with Docker Compose, dropping environment setup from 2h → 5m and making runs reproducible across compute nodes.
• —Designed a hybrid CNN-LSTM network over high-frequency IMU time-series for human-activity / depression-biomarker detection, reaching 87% classification accuracy.

Android (Kotlin) / Docker Compose / CNN-LSTM / IMU Sensor Fusion / Signal Processing

• —Built a time-series sales forecasting engine with XGBoost, engineering lag features to capture seasonality — 92% prediction accuracy drove inventory planning for the Shanghai region.
• —Automated ETL against the data warehouse with CTE-rewritten, index-tuned SQL, cutting extraction latency by 40% for downstream reporting.
• —Replaced per-row loops with Pandas vectorization over high-dimensional transaction data to surface sales trends that informed production strategy.

Python / SQL / XGBoost / Scikit-learn / Pandas / ETL