SERVER Burners

Enhanced Investor Document Verification System: SERVER Burners

Inspiration

The world of venture capital is flooded with pitch decks, and verifying the credibility of these documents is often a manual and error-prone process. We were inspired by the power of AI to automate due diligence, reduce human error, and uncover hidden risks. Tools like large language models and Perplexity Sonar presented a unique opportunity to create a system that could act as a “digital analyst team,” making investment decisions faster, smarter, and more transparent.

What it does

Our system automates the entire investor document verification pipeline:

Accepts PDF/DOCX pitch decks via a web-based dashboard.
Parses and extracts relevant data using NLP.
Engages multiple specialized AI agents (financial, legal, environmental, competitor analysis, claim validation).
Aggregates their findings into a coherent, traceable report.
Outputs insights through visual dashboards and downloadable reports in PDF, JSON, and HTML formats.
Tracks every step with audit logs for compliance.

How we built it

Backend (FastAPI + Celery):

High-performance async APIs for responsiveness.
SQLAlchemy ORM + Pydantic models for robust data handling.
Celery + Redis for background processing.

Frontend (Streamlit):

Real-time status updates, file uploads, and visual report interface.
Score visualization and downloadable reports.

Document Parsing & NLP:

Used PyMuPDF, pdfplumber, python-docx for ingestion.
Applied spaCy and HuggingFace transformers for NER and fact extraction.

External Integration:

Integrated Perplexity Sonar API for real-time web research.
Implemented rate-limiting, retries, and fallbacks for reliability.

DevOps:

Local and Docker-based deployment support.
Auto-setup scripts and documentation for contributors.

Challenges we ran into

Async Orchestration: Coordinating agents and background tasks without race conditions.
API Reliability: Handling Perplexity API rate limits and downtime with retries and graceful fallbacks.
Cross-Platform Dependencies: Issues with libmagic and Redis on Windows required extra setup steps.
Serialization: Ensuring nested models from multiple agents were JSON-serializable.
Schema Syncing: Rapid iteration meant constant database and model updates.
Frontend Feedback: Communicating real-time task progress in Streamlit while handling long processing times.

Accomplishments that we're proud of

End-to-End Automation: Complete pipeline from pitch upload to insight report.
Multi-Agent System: Specialized agents contribute domain-specific analysis.
Explainability: Traceable decisions backed by full audit logs and detailed outputs.
Security & Compliance: GDPR features and robust validation built-in.
Scalable Design: Easily extensible to new agents and deployment environments.

What we learned

How to orchestrate multiple AI agents and combine their outputs effectively.
Async programming and background processing with FastAPI + Celery.
Robust integration of external APIs with retry and fallback logic.
Importance of clean data modeling and serialization with SQLAlchemy and Pydantic.
Best practices for compliance, including audit trails and user data safety.

What's next for SERVER Burners

Agent Expansion: Add more agent types (e.g., fraud detection, market sentiment).
LLM Optimization: Fine-tune smaller models for local inference to reduce API dependency.
Collaboration Features: Enable investors to comment, annotate, and share findings securely.
Enterprise Features: Role-based access, multi-user dashboards, and advanced analytics.
AI Feedback Loop: Train models using feedback from real investment decisions to improve accuracy.