Veritas

Bridging the gap between predictive data and strategic reasoning.

Comment

Inspiration

The startup world is often governed by “gut feelings” and hype, which contributes to an extremely high startup failure rate.

As AIML students, we identified a clear gap: although there is vast historical data about startup success and failure, this information is rarely accessible or understandable to everyday founders.

This motivated us to build Veritas — a platform designed to democratize institutional-grade investment intelligence.

Instead of providing only simple “pass/fail” predictions, Veritas focuses on delivering explainable insights, helping founders understand why a prediction was made and what factors influence success.

What It Does

Veritas is an AI-powered decision-support platform that evaluates a startup’s likelihood of success.

It accepts structured startup inputs such as:

  • Funding stage
  • Market sector
  • Geographic region
  • Startup age
  • Business characteristics

The system performs a two-stage analytical process.

1. Quantitative Prediction

An ensemble of machine learning models, trained on 49,000+ real-world startup records, generates a startup success probability score.

2. Qualitative Reasoning

The prediction score is then passed to Amazon Nova (via AWS Bedrock), which simulates a virtual investment committee that analyzes the startup from multiple perspectives.

The platform runs an Agentic AI workflow that evaluates:

  • Risk factors
  • Market dynamics
  • Competitive positioning
  • Pitch deck content

Finally, the system produces a comprehensive Investor Brief, explaining the startup’s strengths, weaknesses, and potential path to success.

How We Built It

Veritas was developed using a modern production-grade architecture.

Machine Learning Engine

Implemented in Python using:

  • XGBoost
  • Scikit-Learn
  • Pandas

Key design elements include:

  • A custom preprocessing pipeline for high-cardinality categorical features
  • Techniques to address class imbalance in startup outcomes
  • Feature engineering to improve predictive performance

The AI Reasoning Layer

We integrated Amazon Nova (Pro and Lite) via AWS Bedrock to handle:

  • Complex reasoning
  • Multi-turn advisory chat
  • Multimodal pitch deck analysis

This layer acts as an AI investment committee interpreting the machine learning outputs.

Backend Infrastructure

The backend was implemented using FastAPI, providing:

  • High-performance inference endpoints
  • ML model orchestration
  • Integration with Bedrock agent workflows
  • Asynchronous request handling

Frontend Dashboard

The user interface was developed using:

  • React.js
  • GSAP animations
  • A dark-themed data dashboard

The interface provides:

  • Interactive startup analysis
  • Visual probability indicators
  • Investor brief generation
  • Real-time AI advisory interaction

Mathematical Model Reliability

Our best-performing model (XGBoost) was optimized using the ROC-AUC metric, which measures the model’s ability to distinguish between successful and failed startups.

AUC = \int_{0}^{1} TPR(FPR^{-1}(x)) , dx

This ensures that the predicted probabilities are well-calibrated and reliable for decision support.

Challenges We Ran Into

Data Mismatch

Training on 49,000 historical startup records required careful alignment with live user inputs.

Differences in feature representation caused inconsistencies between:

  • training data
  • real-time inference data

We solved this by implementing a strict feature alignment pipeline.

Multimodal Latency

Processing pitch deck images through AI vision models introduced latency challenges.

To address this we implemented:

  • asynchronous API calls
  • optimized Bedrock request handling
  • timeout and retry mechanisms

These optimizations ensured a smooth user experience despite heavy AI processing.

Accomplishments We’re Proud Of

Seamless AI + ML Integration

We successfully bridged the gap between:

  • Black-box ML predictions
  • Transparent AI reasoning

Agentic AI Orchestration

We designed a 4-stage autonomous AI agent workflow capable of:

  • analyzing startup risk
  • evaluating market conditions
  • reviewing pitch material
  • generating an investor-grade brief

In some cases the agent can even challenge the ML model’s prediction based on market context.

System Robustness

Our platform achieved a 94.4% pass rate across the end-to-end validation suite, ensuring stability across the entire AI pipeline.

What We Learned

This project reinforced an important principle:

AI is most powerful when it is interpretable.

A prediction such as “80% chance of success” has limited value on its own. However, a detailed report explaining why that score was produced and how the startup can improve it becomes a practical decision-making tool.

Through this project we gained deep experience in:

  • AWS Bedrock orchestration
  • Agentic AI workflow design
  • Prompt engineering
  • Multimodal AI integration
  • Clean architecture for full-stack AI systems

What's next for Veritas

  • Live Data Integration: Connect with APIs such as Crunchbase and LinkedIn to automatically track real-time startup updates like funding, hiring trends, and market activity.

  • AI Pitch Deck Generator: Use Amazon Nova to help founders improve their pitch decks by identifying weaknesses in their startup profile and suggesting stronger slides, metrics, and storytelling.

  • Collaborative Mode: Introduce a shared workspace where teams of investors can interact with the AI agent together, discuss startup analyses, and generate joint investor insights.

Built With

Share this project:

Updates