Ruoying Wang

Advisor: Prof. Jiacheng Shen

Optimized the draft-proposal path for training-free LLM inference. Multi-source suffix retrieval, trie-based candidate fusion, and confidence-calibrated, cost-aware selection reduced decode-phase wall time by 20.4% against the original Suffix proposer on Llama 3.1-8B with an NVIDIA A40, across five full benchmark runs.

Reworked sequential per-request speculation into batched, cross-request parallel proposal with a persistent C++ thread pool and GIL release. At batch size 16, CPU proposal time fell from 280.6 to 102.9 μs (2.73× speedup) with bitwise-identical drafts.

Advisor: Prof. Jiacheng Shen

Co-designed a compact reduction DSL and confidence-gated memory retrieval to reuse successful GUI reasoning and avoid repeated vision-language inference. On AndroidControl with Qwen3-VL 8B, per-step latency fell from 3,314 to 1,019 ms (3.25× speedup), while low-level step accuracy improved from 65.2% to 69.3% over a chain-of-thought agent. Manuscript submitted to ICLR.

Advisor: Prof. Zuchuan Li

Built a distributed PyTorch training and streaming data pipeline for terabyte-scale satellite imagery, focusing on multi-GPU throughput and I/O efficiency. Developed 256-level spectral tokenization with learnable embeddings for MAE pretraining on 170+ band PACE hyperspectral imagery. Downstream chlorophyll-a estimation reached R² = 0.986, with approximately 2.5× faster convergence from MAE pretraining.

Foshan Xinglun Laboratory Technology Co., Ltd.

Built and launched an AI-driven HKDSE question-bank platform on Moodle LMS for five subjects—English, Mathematics, Physics, Chemistry, and Biology—with approximately 11,800 questions and 625+ quiz sets.

Designed an AI transcription → automated deployment → programmatic approval pipeline for question content. Integrated backend model inference for AI-assisted grading and assessment analysis, connecting question management, marking, analytics, and school-facing delivery.

In a live test, the student-facing answer-entry page measured approximately 0.2 s TTFB.

Nanjing LES Electronic Equipment Co., Ltd.

Built a real-time UAV video inference pipeline combining YOLOv5 detection, Qwen2.5-VL scene understanding, and temporal consistency checks across frames. For deployment on Huawei Ascend 910B NPUs, applied structured pruning, INT8 quantization, and hardware-aware compilation to reduce model footprint and inference overhead.

Engineered an asynchronous five-stream 720p pipeline that overlaps hardware video decoding, preprocessing, and NPU execution. The deployed inference path sustained P95 decode-to-inference latency below 100 ms.