Research Publications

An Automated Startup Evaluation Pipeline: Startup Success Forecasting Framework

Wang, X., & Ihlamur, Y.2024
NAACL Industry Track (Under Review)
Timeline: March 2024 - Oct 2024

Key Highlights

  • Developed RAG-based system for automated startup evaluation
  • Created 3-block multi-agent framework for venture capitalist analysis
  • Improved accuracy by 26% compared to vanilla GPT-4
  • Conducted LLM Founders Segmentation on 2000+ profiles
  • Developed Founder-Idea-Fit Network with 0.0386 validation loss
  • Trained LLM-based Random Forest network
  • Open-sourced codebase and built interactive demo

Impact

  • First-author paper submitted to NAACL Industry Track 2025
  • Collaboration with Vela Partners
  • First-author paper submitted to NAACL Industry Track, 2025
  • Open-sourced the codebase and built an interactive demo showcasing a RAG-based system and Founder-Idea-Fit Network

Agent for Machine Learning: A Text to Model Approach

Wang, X.2024
8th Future of Information & Communication Conference
Timeline: June 2024 - Sept 2024

Key Highlights

  • Developed fully automated system for ML model generation from natural language
  • Automated entire process from data preprocessing to model generation
  • Achieved 0.2130 average Normalized Mean Square Error across 10 Kaggle datasets
  • Implemented Multi-Agent system for comprehensive automation
  • Delivered interpretable results and automated reports

Impact

  • Accepted as first-author paper

None-Line-Of-Sight Imaging Beyond the Corner

Wang, X.2023
Tencent Aspiring Explorers in Science
Timeline: June 2022 - March 2023

Key Highlights

  • Modelled NLOS Imaging with five bounces of diffuse reflections
  • Proposed new algorithm for NLOS image reconstruction
  • Transformed algorithm into executable discrete forms
  • Generalized algorithm to Nth order bounces
  • Achieved good quality in simulation results

Impact

  • National First Place winner
  • Supervised by Prof. Feihu Xu, USTC
  • Selected for program with 5% acceptance rate

White Noise Testing on the LSTM Network Trained with Double Pendulum

Wang, X.2021
5th International Conference on Electrical, Mechanical and Computer Engineering
Timeline: 2021.05-2021.11

Key Highlights

  • Examined computational relations between chaos and randomness
  • Simulated chaotic double pendulum dynamics using Lagrangian equations
  • Engineered LSTM model with MSE of 0.0019
  • Applied Ljung-Box and Box-Pierce tests for randomness assessment
  • Revealed seasonal and chaotic patterns in the data

Impact

  • Top 10 in S.T.-Yau Global Final
  • Published in Journal of Physics Conference Series
  • Oral presentation at 5th ICEMCE
  • Supervised by Prof. Mario Campanelli & Prof. Nicolai Reshetikhin