Using funding-network features and graph ML to predict which US AI startups succeed, with a VC mentor.
Background
A team course project mentored by a venture capital firm. We predicted success for US AI startups founded 2015–2025 using age-tiered funding thresholds (e.g., $5M+ raised within 2 years, $25M+ within 4 years, up to $140M+ within 10 years). Beyond company data such as round sizes and headcount, we modeled relationships between founders, angels, investor-firm founders, and board members through shared universities and prior employers. The team compared logistic regression, Node2vec, XGBoost, LightGBM, random forests, GNN+GAT, and GraphSAGE.
My role
Led data acquisition from the Crunchbase API, building a phased, AI-assisted (Claude Code) pipeline that pulled target-industry companies, applied filters step by step, and validated required columns.
Caught misclassified non-AI companies in the first pull and iteratively refined industry definitions, producing a final dataset of 6,700 companies, 19,000 funding rounds, 16,000+ investors and partners, and 80,000+ people.
Built the XGBoost and GraphSAGE models, training each on tabular-only features and again with engineered network features such as investment degree.
Deliverable
Slide deck (PDF)
The slide deck opens best in your device's PDF viewer.
Results
Startups backed by VCs with large, diverse portfolios succeed more often, and co-investment density matters: strong investors cluster on the same deals.
Bridge investors punch above their weight, while founders' social connections help only modestly.
Limitations: ~50% missing people data, and the model cannot separate VCs causing success from VCs selecting winners.