🚀 VentureFlow Intelligence

AI-Driven Startup Success Prediction Using a Dual-Encoder Fusion Architecture

---

Author	Annanahmed Shaikh
Email	shaikha4@wit.edu
Program	Master of Science in Data Science
University	Wentworth Institute of Technology
Course	DATA-6900: Capstone Project
Semester	Spring 2026
Live Dashboard	Streamlit Cloud

---

� Abstract

Venture capital investors deploy over $300 billion annually into early-stage startups, yet 75–80% of these investments fail to return capital. VentureFlow Intelligence addresses this challenge by introducing a Dual-Encoder Fusion architecture that combines natural language understanding with financial signal processing to predict startup success.

The system processes 47,833 real-world startup records sourced from Crunchbase, spanning 17 years (2005–2022) across 15+ industry sectors. Unlike traditional tabular classifiers, VentureFlow reads and understands the business description of each startup using a pre-trained Sentence-BERT transformer, while simultaneously evaluating financial fundamentals through a log-normalized funding scorer.

Key Result: The Dual-Encoder achieves a ROC-AUC of 0.740 and a Top-10% Precision of 85.3%, outperforming all baseline models by over 10%.

---

🧠 Core Innovation: The Dual-Encoder Fusion Model

Traditional machine learning models treat startup prediction as a simple tabular classification problem. VentureFlow introduces a fundamentally different approach:

Branch A — Semantic Similarity Encoder (The "Reader")

• Uses Sentence-BERT (all-MiniLM-L6-v2) with 22 million parameters
• Converts each startup's text description into a 384-dimensional dense vector
• Compares the vector against a "Unicorn Centroid" — the averaged embedding of all historically successful companies
• Produces a semantic similarity score between 0 and 1

Branch B — Financial Signal Encoder (The "Accountant")

• Applies log-normalization to total funding amount
• Caps extreme values at $1 billion to prevent outlier distortion
• Scales to a 0–1 range
• Acts as a "reality check" — no funding means no high score

Fusion Layer

The two branch scores are combined using an empirically validated weighted formula:

Final Score = 0.60 × Financial Score + 0.40 × Semantic Score

This 60/40 split was determined through an 11-step ablation study and represents the optimal balance between financial reality and semantic innovation.

---

📊 Key Results

Metric	Value	Meaning
ROC-AUC	0.740	The model reliably distinguishes winners from failures (0.50 = random guess, 1.00 = perfect)
Top-10% Precision	85.3%	When the model picks its top 10% most confident predictions, 85 out of 100 are actual winners
Improvement over Baseline	+10%	The Dual-Encoder outperforms the best traditional model (Random Forest at 0.675)

Model Comparison Table

Model	Type	ROC-AUC	Notes
Logistic Regression	Tabular only	0.565	Near coin flip
Random Forest	Tabular only	0.675	Best traditional baseline
Early Fusion (RF + LLM)	Naive concatenation	0.642	Adding text hurt performance
Dual-Encoder (Ours)	Intermediate fusion	0.740	Best performer — 10% improvement

Critical Finding: Naively concatenating text embeddings with financial features (Early Fusion) actually degraded performance from 0.675 to 0.642. This validates our architectural decision to keep the two modalities separate until the final scoring phase.

---

📡 Live Data Validation (Out-of-Distribution Testing)

To test temporal generalization, we scored 6 prominent 2024–2025 AI companies that were not present in the training data:

Startup	Sector	Total Funding	AI Score
xAI	Generative AI	$6.0B	0.990
Anduril Industries	Defense Tech	$1.5B	0.924
Safe Superintelligence	AGI Research	$1.0B	0.881
Figure AI	Robotics	$675M	0.853
Mistral AI	Open Source LLM	$640M	0.847
Anthropic	AI Safety	$7.3B	0.658

Insight: Anthropic scored lower (0.658) despite having the most funding ($7.3B) because its description focuses on "AI safety" — a concept absent from historical unicorn language. This demonstrates that the model reads meaning, not just money.

---

🗄️ Database Architecture: Star Schema

The cleaned data is stored in an SQLite database organized as a Star Schema — the same dimensional modeling approach used by Amazon, Netflix, and Spotify for analytics.

┌─────────────────────┐     ┌──────────────────────────┐     ┌──────────────────┐
│   dim_startup       │     │   fact_funding_rounds     │     │   dim_time       │
│─────────────────────│     │──────────────────────────│     │──────────────────│
│ startup_id (PK)     │────▶│ round_id (PK)            │◀────│ date_key (PK)    │
│ name                │     │ startup_id (FK)           │     │ year             │
│ category_list       │     │ funding_round_type        │     │ quarter          │
│ status              │     │ funding_round_code        │     │ month            │
│ country_code        │     │ raised_amount_usd         │     └──────────────────┘
│ state_code          │     │ participants              │
│ city                │     │ is_first_round            │
│ founded_at          │     │ is_last_round             │
│ first_funding_at    │     └──────────────────────────┘
│ last_funding_at     │
│ description         │
│ embeddings_json     │
└─────────────────────┘

---

📁 Repository Structure

Investment_Intelligence_Platform/
│
├── Code/                          # All source code
│   ├── app/
│   │   └── dashboard.py           # Streamlit dashboard (backup copy)
│   ├── models/
│   │   ├── deep_llm_fusion.py     # Dual-Encoder Fusion model class
│   │   └── deep_llm_v1.pkl        # Trained model weights
│   ├── scripts/
│   │   ├── etl_pipeline.py        # Data cleaning & loading pipeline
│   │   ├── fetch_live_data.py     # Live API data integration (NewsAPI + Companies API)
│   │   ├── ablation_study.py      # 11-step weight sensitivity analysis
│   │   ├── generate_embeddings.py # Sentence-BERT embedding generator
│   │   ├── generate_report_graphs.py        # Thesis visualization generator
│   │   ├── generate_thesis_performance_report.py  # Full performance report
│   │   ├── hypothesis_test_q1_2026.py       # Statistical hypothesis testing
│   │   ├── import_new_data.py     # New startup data importer
│   │   ├── benchmark_live_impact.py         # Live vs. historical benchmarking
│   │   ├── show_updates.py        # Database update viewer
│   │   ├── verify_db.py           # Database integrity checker
│   │   └── verify_report.py       # Report output validator
│   ├── dashboard.py               # Main Streamlit dashboard (664 lines)
│   ├── deep_llm_fusion.py         # Model definition (root copy)
│   └── requirements.txt           # Python dependencies
│
├── Datasets/                      # Data files
│   ├── startup_data_cleaned.xlsx  # Cleaned dataset (N=47,833)
│   └── schema.sql                 # Star Schema DDL (3 tables)
│
├── Visualization/                 # All generated graphs and reports
│   ├── thesis_report/
│   │   ├── model_comparison.png          # 4-model ROC-AUC bar chart
│   │   ├── performance_evolution.png     # Baseline → Fusion improvement timeline
│   │   ├── semantic_discovery_map.png    # t-SNE startup embedding clusters
│   │   ├── sector_velocity.png           # Sector-level success velocity
│   │   ├── investment_opportunity.png    # Risk-vs-reward investment matrix
│   │   ├── live_data_impact.png          # Live company score distribution
│   │   ├── capital_inefficiency.png      # Funding inefficiency heatmap
│   │   ├── persona_clustering.png        # Investor persona segmentation
│   │   ├── flowchart of project.png      # End-to-end system workflow diagram
│   │   └── star_schema_diagram.png       # Database ERD diagram
│   ├── ablation/
│   │   ├── ablation_curve.png            # Weight sensitivity curve
│   │   └── ablation_results.csv          # Raw ablation data (11 steps)
│   └── hypothesis_2026/                  # Statistical test outputs
│
├── Other/                         # Academic deliverables
│   ├── Final_Thesis_Presentation_v2.pptx  # 18-slide defense presentation
│   ├── Data_6900_P8_Final_Thesis_Annanahmed_Shaikh.pdf  # Thesis manuscript (PDF)
│   └── FINAL_30PAGE_OVERLEAF_THESIS.tex   # LaTeX source (30 pages, 1200+ lines)
│
└── README.md                      # This file

---

⚙️ Technology Stack

Category	Technology	Purpose
Language	Python 3.10+	Core implementation
NLP Model	Sentence-BERT (all-MiniLM-L6-v2)	Text embedding generation
ML Framework	scikit-learn	Baseline models & metrics
Database	SQLite	Star Schema data warehouse
Dashboard	Streamlit	Interactive web application
Visualization	Matplotlib, Seaborn, Plotly	Graphs and charts
Data Processing	Pandas, NumPy	ETL pipeline
Live Data	NewsAPI, The Companies API	Real-time startup discovery
Deployment	Streamlit Cloud	Production hosting

---

� Quick Start Guide

Prerequisites

• Python 3.10 or higher
• pip package manager

Installation

# Clone the repository
git clone https://github.com/veagy/Investment_Intelligence_Platform.git
cd Investment_Intelligence_Platform

# Install dependencies
pip install -r Code/requirements.txt

# Launch the dashboard
streamlit run Code/dashboard.py

Environment Variables (Optional — for Live Data)

Create a .env file in the project root:

NEWS_API_KEY=your_newsapi_key_here
COMPANIES_API_KEY=your_companies_api_key_here

Note: The system works without API keys by falling back to a curated list of 2024–2025 AI companies.

---

🔬 Methodology Summary

1. Data Collection: 47,833 startup records from Crunchbase (2005–2022)
2. Data Cleaning: Automated ETL pipeline with outlier capping, text normalization, and temporal validation
3. Data Warehousing: SQLite Star Schema (dim_startup, dim_time, fact_funding_rounds)
4. Baseline Models: Logistic Regression (0.565) and Random Forest (0.675)
5. Dual-Encoder Design: Parallel processing of text (Branch A) and financials (Branch B)
6. Ablation Study: 11-step weight sweep from 0% to 100% financial weight
7. Optimal Configuration: 60% Financial + 40% Semantic = 0.740 ROC-AUC
8. Live Validation: Out-of-distribution testing on 6 unseen 2024–2025 companies
9. Deployment: Production Streamlit dashboard with 8 interactive tabs

---

� References (APA Format)

• Arroyo, J., Corea, F., Jimenez-Diaz, G., & Recio-Garcia, J. A. (2019). Assessment of machine learning performance for decision support in venture capital investments. IEEE Access, 7, 124233–124243.
• Baltrusaitis, T., Ahuja, C., & Morency, L.-P. (2019). Multimodal machine learning: A survey and taxonomy. IEEE TPAMI, 41(2), 423–443. https://doi.org/10.1109/TPAMI.2018.2798607
• Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5–32. https://doi.org/10.1023/A:1010933404324
• Chawla, N. V., Bowyer, K. W., Hall, L. O., & Kegelmeyer, W. P. (2002). SMOTE: Synthetic minority over-sampling technique. JAIR, 16, 321–357. https://doi.org/10.1613/jair.953
• Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019). BERT: Pre-training of deep bidirectional transformers. In NAACL-HLT (pp. 4171–4186). https://doi.org/10.18653/v1/N19-1423
• Gompers, P., & Lerner, J. (2001). The venture capital revolution. Journal of Economic Perspectives, 15(2), 145–168. https://doi.org/10.1257/jep.15.2.145
• Kimball, R., & Ross, M. (2013). The Data Warehouse Toolkit (3rd ed.). Wiley.
• Reimers, N., & Gurevych, I. (2019). Sentence-BERT: Sentence embeddings using Siamese BERT-Networks. In EMNLP-IJCNLP (pp. 3982–3992). https://doi.org/10.18653/v1/D19-1410
• Vaswani, A., et al. (2017). Attention is all you need. In NeurIPS (pp. 5998–6008).
• Żbikowski, K., & Antosiuk, P. (2021). A machine learning, bias-free approach for predicting business success using Crunchbase data. Information Processing & Management, 58(4), 102555.

---

📄 License

This project was developed as part of the Master of Science in Data Science program at Wentworth Institute of Technology. All rights reserved.

---