LaunchMintAI: Forensic Startup Intelligence & Quantitative Validation Platform 🚀

Applied Data Science & Machine Learning Platform
Validating Day-0 Pre-Seed Venture Concepts across 189,970 Historical Venture Cohorts with Leak-Free Predictive Modeling, Stochastic Differential Equations, and Deterministic RAG Triad Grounding.

---

📌 Executive Summary

Early-stage venture validation is fundamentally broken: over 90% of tech startups fail within 36 months, yet founders and angel investors routinely rely on subjective bias or generic LLM "roasters" that hallucinate market sizes and invent survival numbers.

LaunchMintAI replaces ungrounded generative advice with an empirical, multi-disciplinary Applied Data Science and Quantitative Simulation Architecture:

1. Predictive Machine Learning: Regularized XGBoost survival classifier trained on 189,970 historical Crunchbase startups using strictly Day-0 pre-seed observable features (Holdout ROC-AUC: 0.8512, PR-AUC: 0.4789).
2. Quantitative Financial Modeling: Vectorized 10,000-iteration NumPy Monte Carlo engine calculating path-dependent cash flow bounds, runway ruin probability $P(\text{ruin})$, and 95% Value at Risk (VaR) in $< 32\text{ms}$.
3. Aspect-Based Sentiment NLP: Tokenized VADER Aspect NLP extracting customer friction vectors (Pricing, Product Reliability, Support Latency) across competitor reviews to compute Competitor Vulnerability Indices (CVI).
4. Deterministic Retrieval & Grounding: 3-tier domain authority search waterfall paired with regex fact-checking, achieving 95.8% Faithfulness and 0.0% Hallucination Rate across 30 golden evaluation prompts.
5. Tactical UI Integration: React 19 / Vite 6 tactical terminal connected via FastAPI to render real-time calibrated survival percentages and live SHAP TreeExplainer attributions.

---

🏛️ 5-Pillar System Architecture

                                ┌──────────────────────────────────────────────────────────┐
                                │                 LaunchMintAI DS ENGINE                   │
                                └──────────────────────────────────────────────────────────┘
                                                              │
          ┌──────────────────────────┬────────────────────────┴───────────────────────┬──────────────────────────┐
          ▼                          ▼                                                ▼                          ▼
    [PILLAR 1: ML]        [PILLAR 2: QUANT/STATS]                              [PILLAR 3: NLP]           [PILLAR 4: RETRIEVAL]
   XGBoost Survival          10,000-Run Monte Carlo                               Competitor Pain            3-Tier Grounded
      Classifier               Financial Simulation                              Sentiment Analysis         Waterfall Search
    (Day-0 Gated)                  (Vectorized)                                     (Aspect VADER)              (RAG Triad)
          │                              │                                                │                          │
Trained on 189,970             Runs 10,000 parallel                             Scrapes competitor reviews  Tier-1 research domains
historical startups.           paths in vectorized NumPy:                       and scores customer pain   (Statista/Gartner) +
Outputs survival prob %        - Bear/Base/Bull cash flows                      density across pricing,    Regex Grounding Gate +
+ SHAP explainability.         - Runway $P(\text{ruin})$ & 95% VaR              support, and uptime.       Ragas Faithfulness SLA.

---

🔬 Forensic Case Study: V1 Naive Prototype vs. V2 Production Engine

In early prototype iterations, our baseline model hit an apparent 0.9249 ROC-AUC. An adversarial Applied Data Science audit diagnosed fatal Target Definition Leakage and Temporal Incoherence:

• In Crunchbase, is_success was defined as acquired | ipo | (operating & (funding >= $5M | rounds >= 3)).
• Supplying cumulative funding_total_usd and funding_rounds allowed decision trees to memorize the target definition rather than learning genuine early venture signals.
• Furthermore, asking pre-seed founders on Day 0 for downstream funding variables created massive serving-training skew.

We executed a complete architectural refactor, purging all post-outcome features and restricting the model strictly to Day-0 observables:

Metric / Property	V1 Naive Prototype (Pre-Remediation)	V2 Production Engine (Current State)	Applied Data Science Rationale
Feature Space Horizon	10-Year Post-Outcome Variables (funding_total_usd, rounds, milestones)	Strictly Day-0 Pre-Seed Observables (founder_team_size, is_tier_1_hub, competitor_density, 12 verticals)	Zero Target Leakage. Aligns training horizon with raw pre-seed validation prompts.
Target Leakage Status	Fatal Leakage Present	Clean (0 Leakage)	V1 memorized the target formula; V2 discovers genuine causal signals.
5-Fold CV ROC-AUC	0.9199 ± 0.0012	0.8497 ± 0.0017	Negligible variance across 5 folds proves model stability without overfitting.
Holdout Test ROC-AUC ($N=37,994$)	0.9249 (Cheated / Memorized)	0.8512 (Real Generalization)	An ROC-AUC of 0.8512 on Day-0 signals alone is an elite result in venture capital economics.
Holdout Test PR-AUC	0.7630 (Artificially Inflated)	0.4789 (Defensible)	In a 9:1 imbalanced domain (base rate 9.89%), 0.4789 represents a ~5x precision lift over random baseline.
Brier Calibration Score	0.0872	0.1562	Well-calibrated, monotonically increasing probabilities across 6 risk tiers.
Optimal F1 Score	0.7444 (@ $\tau = 0.825$)	0.4286 (@ $\tau = 0.600$)	Balanced operational threshold under severe imbalance.
RAG Evaluation Method	Simulated np.random.uniform()	Deterministic Regex Grounding & Authority Engine	True empirical verification across 30 golden test prompts in 11 startup verticals.
RAG Faithfulness / Groundedness	Simulated 97.8%	95.8% (vs Baseline 66.4%)	+29.3% measured groundedness uplift over unanchored zero-shot LLM.
RAG Context Precision	Simulated 94.7%	91.8% (vs Baseline 56.2%)	+35.6% domain authority precision from Tier-1 waterfall search.
RAG Hallucination Rate	Simulated 0.0%	0.0% (vs Baseline 33.3%)	Zero fabricated numerical claims on adversarial concepts.
Serving Latency (P95)	2,217 ms	385 ms	-73.4% latency reduction via vectorized NumPy math and decoupled LLM generation.
Frontend/ML Coupling	Disconnected (LLM hallucinated % score)	Coupled via Live FastAPI Bridge	React client queries live XGBoost classifier and displays exact SHAP drivers.

---

📊 Exploratory Data Analysis & Empirical Evidence

The Crunchbase venture dataset (189,970 startups founded 1995–2014) reveals profound structural venture dynamics:

Metric	Empirical Value	Applied Data Science Significance
Total Cohort Size	189,970	Minimum 10-year observation window for terminal outcomes.
Positive Class ($Y=1$)	18,793 (9.89%)	Startups achieving verified acquisition, IPO, or sustainable velocity.
Negative Class ($Y=0$)	171,177 (90.11%)	Startups closed, defunct, or stranded.
Class Imbalance Ratio	9.11 : 1	Optimized via scale_pos_weight = 9.11 without SMOTE distortion.
Top Sector Survival Rate	37.41% (HealthTech & Bio)	High-CapEx / IP moats exhibit significantly higher acquisition rates.
Lowest Sector Survival Rate	5.81% (EdTech)	Low barriers to entry create hyper-saturated cohorts with high attrition.
Tier-1 Tech Hub Uplift	+11.4% Baseline	SF, NYC, Boston, London, Tel Aviv provide persistent geographic tailwinds.

All generated visual artifacts are stored in backend/data/eda_plots/:

• 01_class_imbalance.png: Empirical 9.11:1 startup mortality distribution.
• 02_survival_by_macro_vertical.png: Structural survival variances across 12 sectors.
• 03_funding_distribution_kde.png: Bimodal capital separation.
• 04_correlation_heatmap.png: Inter-feature covariance matrix.
• 05_shap_feature_importance.png: Global SHAP TreeExplainer feature attributions.
• 06_rag_triad_benchmark.png: Deterministic RAG Triad evaluation metrics.

---

⚙️ Technical Stack

• Machine Learning: xgboost, shap (TreeExplainer), scikit-learn, joblib.
• Quantitative Simulation: numpy (Vectorized SIMD broadcasting, Gaussian stochastic processes).
• NLP: nltk (VADER Aspect-Based Sentiment Analysis).
• Backend API: fastapi, uvicorn, pydantic, sqlmodel, chromadb, slowapi.
• Frontend Client: react 19, typescript, vite 6, tailwindcss, lucide-react.
• Retrieval & Research: tavily-python, google-generativeai.

---

🔌 API Reference & Serving Schemas

1. Day-0 Survival Prediction & Live SHAP Attribution

POST /predict_survival
Content-Type: application/json

{
  "macro_vertical": "SaaS & Enterprise",
  "founder_team_size": 4,
  "is_tier_1_hub": 1,
  "competitor_cohort_density": 850
}

Response (in $<15\text{ms}$):

{
  "status": "success",
  "macro_vertical": "SaaS & Enterprise",
  "survival_probability": 0.4682,
  "survival_percentage": "46.8%",
  "risk_tier": "MODERATE_RISK",
  "sector_baseline": "14.7%",
  "delta_vs_baseline": "+32.1%",
  "shap_drivers": {
    "positive_factors": [
      "+18.4% · Core Team & Founder Size",
      "+12.1% · Tier 1 Venture Hub Location"
    ],
    "risk_factors": [
      "-4.2% · Sector Cohort Density"
    ]
  },
  "model_metadata": {
    "algorithm": "XGBoost Classifier + SHAP TreeExplainer (Day-0 Leak-Free)",
    "training_samples": 151976,
    "test_roc_auc": 0.8512,
    "test_pr_auc": 0.4789
  }
}

2. 10,000-Iteration Vectorized Monte Carlo Simulation

POST /simulate_financials
Content-Type: application/json

{
  "macro_vertical": "SaaS & Enterprise",
  "initial_capital_usd": 2500000.0,
  "monthly_burn_rate_usd": 60000.0,
  "initial_monthly_revenue_usd": 20000.0,
  "simulation_months": 36,
  "num_simulations": 10000
}

3. Empirical RAG Triad Metrics

GET /eval_metrics

---

📂 Repository Structure

LaunchMintAI/
├── backend/
│   ├── app/
│   │   ├── main.py                     # FastAPI application gateway & REST endpoints
│   │   ├── services/
│   │   │   ├── survival_engine.py      # Day-0 XGBoost survival classifier & SHAP explainer
│   │   │   ├── monte_carlo_engine.py   # Vectorized 10,000-run NumPy cash simulation (<32ms)
│   │   │   ├── sentiment_nlp_engine.py # VADER Aspect-Based Sentiment Analysis (Pricing/Uptime/Support)
│   │   │   ├── market_search.py        # 3-Tier domain authority waterfall search (Tavily)
│   │   │   └── evidence_model.py       # Deterministic regex fact extraction & credibility scoring
│   │   └── models/artifacts/           # Serialized XGBoost model bundle (.joblib)
│   ├── data/
│   │   ├── eda_plots/                  # 6 empirical figures (ROC-AUC, SHAP, Monte Carlo, RAG Triad)
│   │   └── processed/                  # Serialized benchmark outputs (eval_benchmark_results.json)
│   └── scripts/
│       ├── models/                     # Day-0 model training & validation pipeline
│       └── eval/                       # Deterministic RAG Triad benchmark suite
├── frontend/
│   ├── features/                       # Tactical UI modules (VC Roast, War Room, Validator, Pitch Forge)
│   └── services/geminiService.ts       # Frontend API bridge & multi-agent routing
├── docs/
│   ├── ARCHITECTURE_SPEC.md            # Exhaustive 5-layer engineering & mathematical specification
│   └── README.md                       # Developer setup & local execution guide
└── LAUNCHMINT_DOCUMENTATION.md         # Applied DS Case Study, Target Leakage Post-Mortem & Defense Guide

---

🚀 Quickstart & Local Setup

1. Prerequisites

• Python: 3.10+
• Node.js: 18+ (npm 9+)

2. Backend Installation

cd backend
python -m venv venv
# Windows:
.\venv\Scripts\Activate.ps1
# Linux / macOS:
source venv/bin/activate

pip install -r requirements.txt

Create backend/.env:

GEMINI_API_KEY=your_gemini_key
TAVILY_API_KEY=your_tavily_key

Run FastAPI Server:

python -m uvicorn app.main:app --host 0.0.0.0 --port 8000 --reload

3. Frontend Installation

cd frontend
npm install
npm run dev

Client runs at http://localhost:5173.

4. Run Automated Verification Suite

# Verify Day-0 XGBoost Model & SHAP Inference
python -m scripts.eval.test_survival_endpoint

# Verify 10,000-Run Vectorized Monte Carlo Engine (<32ms)
python -m scripts.eval.test_monte_carlo

# Run Deterministic RAG Triad Evaluation Benchmark (30 golden prompts)
python -m scripts.eval.rag_triad_benchmark

---

📚 Technical Documentation

• LAUNCHMINT_DOCUMENTATION.md: Applied Data Science forensic post-mortem, target leakage analysis, and technical defense guide for Senior Applied DS interviews.
• docs/ARCHITECTURE_SPEC.md: Comprehensive 5-layer system and data science engineering specification.
• docs/README.md: Quickstart setup and developer environment guide.

---

📄 License

MIT License. Engineered for forensic startup validation and data science excellence.