Jun 2025 – Nov 2025 · 5 mos · Houston, Texas, United States · On-site

• Design and deploy an MCP-based AI agent integrated with Entrade® to automate trade capture, market data ingestion, and real-time portfolio risk analytics; build LangChain workflows with tool orchestration, session memory, structured JSON outputs, and guarded retries for forward exposure, hedge optimization, and P&L reports.
• Build a production RAG pipeline for desk workflows using FAISS indices on desk docs and market reports; engineer few-shot prompt templates for risk management to improve retrieval precision and model performance.
• Deploy and fine-tune LLaMA for other AI features, including trade scheduling line extraction and parcel match.

Jan 2025 – Jul 2025 · 6 mos · San Diego, California, United States · On-site

• 🧠 NeurIPS 2025 — Towards General Continuous Memory for Vision-Language Models
• Proposed CoMEM, a plug-and-play continuous memory framework for Vision-Language Models (VLMs), that encodes multimodal and multilingual knowledge into 8-token compressed embeddings for efficient reasoning.
• Trained a continuous memory encoder by LoRA fine-tuning a Qwen2.5-VL and a shared-parameter Q-Former, updating only 1.2% of weights on a 15.6K self-synthesized VQA dataset, achieving training efficiency on single H100.
• Surpassed state-of-the-art RAG methods by 7%+ on six knowledge-intensive VQA benchmarks (OK-VQA, A-OKVQA, InfoSeek, etc.) and demonstrated robust cross-lingual generalization in low-resource settings.
• Enabled cross-modal memory transfer to non-vision LLMs, improving reasoning without altering their architecture.

Feb 2024 – Jun 2024 · 4 mos · Beijing, China · On-site

• Developed a multimodal time-series prediction framework combining financial news headlines and historical prices via time-heterogeneous GNNs and LLMs to forecast ranked daily stock returns on the FNSPID Dataset.
• Constructed dynamic inter-company relation graphs by integrating 20-day historical price correlations with LLM-extracted sentiment scores from 500 daily news headlines, and optimized a weighted regression–rank loss function to improve both return estimation and ranking accuracy for long/short portfolio construction.
• Delivered a consistent +4% gain in ACC/ARR over baselines, achieving a Sharpe Ratio of 1.34 on FNSPID backtest.

Feb 2024 – Apr 2024 · 2 mos · Beijing, China

• Macroeconomic analysis, policy evaluation.

May 2023 – Aug 2023 · 3 mos · Singapore, Singapore · On-site

• Developed a scalable rendering and web display pipeline for the Universal Differentiable Simulator, which enabled efficient visualization of complex simulations for rigid bodies, fluids, and cloth dynamics.
• Streamlined workflow by integrating Alembic to cache simulation results and incorporating rendering logic, optimizing storage footprint and retrieval speed by 30%.