Wenfei Liu

A human-in-the-loop market simulation for testing how trading decisions hold up against fundamentals and crowd noise.

• Designing on top of TwinMarket (NeurIPS 2025), injecting real decisions into an LLM-agent-driven market to test resistance to herding and rumours.
• Building a FastAPI bridge between the simulation engine and an interactive Convex + PixiJS front end.
• Scoring decisions against fundamentals and crowd behaviour using Qlib risk-adjusted return metrics.

• Built an n8n workflow that fetches public competitor sources, cleans and hashes content, detects changes deterministically, and uses a bounded LLM stage to generate structured email briefs.
• Added evidence validation, human-review and failure routing, snapshot state, and duplicate-delivery suppression for an inspectable, reliable workflow.

• Benchmarked six models for fraud detection; the feed-forward neural network reached PR-AUC 0.766.
• Diagnosed feature drift with PSI and tested time-weighted loss, sliding-window retraining, and a score-level ensemble.
• Simulated federated learning across three clients, staying within 2% of centralised PR-AUC.
• Built a simulated MLOps layer with promotion gates, a drift dashboard, and a canary rollout tested with McNemar’s test.

• Fine-tuned DistilBERT for complaint classification on Amazon reviews (Macro F1 0.931, ROC-AUC 0.989), with a Hugging Face preprocessing pipeline and named entity recognition for product and brand mentions.