fanmo

Built a 3-layer AI framework for crypto asset allocation, combining knowledge graphs, causal inference, and a multi-agent system for autonomous trading decisions. Designed an LLM-based data pipeline that automatically extracted 305,360 entities and 19 relation types from news, on-chain data, and regulatory filings. Reduced graph noise by 92% using Multi-Head Graph Attention (GAT), keeping only ∼12 critical market connections out of 156 to improve signal quality. Deployed a multi-agent architecture where specialized analyst agents synthesize conflicting crypto signals and execute Buy/Sell/Hold decisions with feedback-driven self-improvement.

Built a production-grade multi-asset alerting system scanning 593 assets per 22.65s cycle, achieving 66.7%–68.9% backtest hit rates on model-qualified 30m/1d signals. Automated full market screening across equities, ETFs, futures, and A-shares on 5m/30m/2h/1d timeframes, replacing manual chart monitoring with a single end-to-end pipeline. Designed a real-time signal platform detecting Elliott-wave structures, generating ML-ready features, ranking opportunities with XGBoost, and delivering trader-ready alerts with chart context via Telegram. Engineered the full stack in Python with multithreaded ingestion, Optuna hyperparameter tuning, automated charting, cooldown control, and signal archiving using yfinance and Alpha Vantage data feeds.

Won First Prize in the Guangdong Computer Design Competition; delivered an AI platform validated for innovative fine-tuning methodologies and high-fidelity historical scene generation. Architected a full-stack AI platform integrating ChatGLM3 and Stable Diffusion with modular components for 'Historical Search Cloud' and 'AI Painting', optimizing generative AI workflows for digital humanities.

Led end-to-end product delivery of an internal digital transformation initiative, achieving 98% UAT satisfaction across all departments. Defined product requirements and drove the migration of 1,000+ personnel and operational records, coordinating cross-functional stakeholders to ensure zero data loss. Redesigned document management workflows through user research and process mapping, cutting cross-departmental retrieval time by 70%.

Delivered a highly-scalable attendance system receiving a near-perfect academic score and formal commendation for technical innovation; independently implemented over 60% of the production-level codebase. Architected a full-stack intelligent attendance system using Django and Vue.js with modular decoupled architecture, integrating real-time biometric processing with administrative workflows. Streamlined the system lifecycle by designing optimized MySQL relational schemas and building an interactive dashboard for real-time data visualization and automated PDF report generation.