Shengyuan Liu

Download CV (PDF)

Education

New York University

New York, US

Master of Science in Data Science

Sep 2026 – May 2028

  • Relevant Courses: Introduction to Data Science, Machine Learning, Programming for Data Science, Data Engineering, Advanced Python.

Imperial College London

London, UK

Bachelor of Science in Mathematics, First Class Honours

Sep 2023 – Jun 2026

  • Relevant Courses: Data Science, Statistical Learning, Optimisation, Time Series, Computational Linear Algebra, Game Theory, Stochastic Process, SDE in Financial Modelling.

Professional Experience

Institute of Automation, Chinese Academy of Sciences

Oct 2025 – Dec 2025

Research Assistant

  • Reviewed video multimodal reasoning and GRPO variants to study adaptive rewards. Ran original TinyLLaVA-Video-R1 training and benchmark evaluations following the paper’s setup.
  • Prototyped GRPO-LEAD-inspired modifications using within-group answer accuracy to estimate question difficulty, difficulty-aware length rewards and advantage reweighting, then initiated modified-model training.

Neusoft Group

Jul 2025 – Sep 2025

Software Engineer

  • Processed 10K+ CheXpert Plus frontal X-rays with complete findings and structured reports. Adapted MedGemma-4B for Chinese reporting generation using 4-bit quantization, LoRA and two A100 GPUs.
  • Built a Deepseek evaluation pipeline measuring disease agreement, omissions, false positives and descriptive accuracy. Fine-tuned scores rose from 40+ to 80/100 against baseline.
  • Built an OCR, de-identification and regex extraction pipeline to structure reports from 2K private mammography cases, enabling MedGemma fine-tuning for BI-RADS prediction.

Projects

KnowledgeBase

Sep 2026 – Present

Personal Project: RAG, FastAPI, React, Chroma, SQLite

  • Built a citation-grounded RAG pipeline with Docling parsing and structural chunking, preserving source spans, table headers, and page references; enabled original-page citation navigation and retained historical evidence after source deletion.
  • Integrated BM25, dense retrieval, RRF, BGE reranking, and parent–child chunking with LLM query decomposition and multihop retrieval, improved annotated evidence character coverage from 90.58% to 99.96% and full-span recall from 59.38% to 68.75% across 16 questions within a 4,096-token evidence budget.
  • Built reproducible evaluations for retrieval, citations, abstention, latency, and token usage, matched PDF regression passes 6/6, with 13/13 expected facts supported and cited.

Kaggle Pokemon TCG AI Battle

Jun 2026 – Aug 2026

Silver Medal, 52 / 6,807 teams

  • Built a PyTorch agent via behavior cloning of public game-play and PPO. Early PPO beat its cloning parent in 52.8% of 6,000 seat-balanced local games.
  • Designed state and action encoders with shared card embeddings, a shared Pokemon object encoder and permutation-invariant pooling for variable legal actions. Distilled 10 matchup specialists via cross-entropy into an 8M-parameter policy.
  • Profiled CPU inference and benchmarked ONNX Runtime against PyTorch, reducing model-forward latency by 44%, from 3.45 ms to 1.93 ms, in single-threaded local tests.

Forecast Combination Puzzle

May 2025 – Jun 2025

Research Project: Time Series Analysis, Random Forest, Boosting

  • Implemented exponential smoothing, Holt–Winters, ARIMA, Random Forest and Boosting models in Python on an M4 daily time series; evaluated forecast accuracy using MSE, RMSE, and MAE.
  • Compared equal-weight, error-based, and covariance-based forecast combinations, found simple averaging outperformed error-based weighting and analyzed how residual correlations affected combination performance.

Technical Skills

Programming: Python (PyTorch, NumPy, Pandas, Matplotlib, FastAPI), TypeScript/JavaScript, HTML & CSS, Go

Tools: PostgreSQL, Git, Linux, Shell, Codex, Claude Code.