## Inspiration

The journey from "I have an idea" to "I have a business" is often stalled by the paralysis of manual research. Founders spend weeks toggling between Google Trends, Yahoo Finance, and persona spreadsheets. We were inspired to build a "Venture Capitalist in a Box"—a system that doesn't just give you a chat interface, but actually executes the multi-phase analysis required for a professional product launch. We wanted to build something that feels like a war room where every expert (Market, Finance, Risk, GTM) is working specifically for you.

## What it does

Venture Forge is a multi-agent orchestration platform. It takes a simple product description and triggers a 4-phase simulation:

  1. Validation: Market trends and competitive SWOT analysis.
  2. Financials: Revenue projections and tiered pricing models.
  3. GTM: Marketing channel strategies and product roadmaps.
  4. Decision: A final executive verdict (GO / NO-GO) with a confidence score.

The system features a Self-Learning Loop where agents are calibrated based on "market feedback" after every run, ensuring the system becomes more reliable over time.

## How we built it

We architected the system using a Master Orchestrator pattern built on Python 3.12 and Flask.

  • Agents: Built using a modular BaseAgent class, leveraging Claude 3.5 Sonnet (via Airia) for high-reasoning tasks.
  • Orchestration: Used asyncio to run validation agents in parallel, significantly reducing the simulation time from minutes to seconds.
  • Self-Learning: Implemented a credibility calibration system using an Exponential Moving Average (EMA) formula: $$S_t = \alpha \cdot (1 - \epsilon) + (1 - \alpha) \cdot S_{t-1}$$ Where $S_t$ is the new credibility score, $\alpha$ is the smoothing factor (0.1), and $\epsilon$ is the normalized error from the simulation feedback.
  • Frontend: A premium dark-mode dashboard built with Vanilla JS and Chart.js for real-time data visualization.

## Challenges we ran into

  • JSON Reliability: Getting LLMs to strictly output valid JSON for the frontend to parse was a battle. We solved this with rigorous system prompting and a custom "retry/repair" loop in the BaseAgent.
  • State Management: Passing the context of 9 different agents through 4 phases without "forgetting" critical details required a nested state object managed by the Orchestrator.
  • Audio Streaming: Implementing real-time Speech-to-Text via Modulate required managing binary WebSockets and ensuring the frontend could handle live transcription updates.

## Accomplishments that we're proud of

  • The Learning Tab: Seeing the live chart of agent credibility scores update and persist across sessions feels like the system is truly "alive."
  • The Flow Diagram: Successfully visualizing a complex 9-agent workflow within the UI to keep users informed of what's happening under the hood.
  • Speed: Getting a full-scale market and financial report in under 20 seconds.

## What we learned

We learned that the true power of AI isn't in a single large prompt, but in Agentic Workflows. Breaking a complex task (like a product launch) into 9 granular steps produced far higher quality results than asking a single LLM to "write a business plan." We also deepened our understanding of WebSocket proxies and the nuances of live financial data integration.

## What's next for VentureForge

  • Multi-Model Support: Letting users choose between different LLM backends (Llama 3, Gemini, etc.) for different agents.
  • Real-World Pivot: Connecting the "Simulation" feedback to actual real-world market performance via APIs like Stripe or Google Analytics.
  • Collaborative Mode: Allowing multiple team members to chat with the Master Agent simultaneously.

Built With

  • airia
  • asyncio
  • braintrust
  • chart.js
  • claude-3.5-sonnet
  • css3
  • flask
  • google-trends
  • html5
  • modulate-velma-api
  • python-3.12
  • pytrends
  • uv
  • vanilla-js
  • yfinance
Share this project:

Updates