BullRunAI

Inspiration

The cryptocurrency market’s 24/7 volatility, fragmented data sources, and unreliable predictions create chaos for traders. Traditional tools fail to adapt to crypto’s rapid shifts, leaving investors drowning in news feeds and static charts. We built BullRun AI to unify 3 AI Agents: Live Real-time data, AI-driven Crypto Price forecasting (powered by Hybrid LSTM with accuracy of 90%+), and LLM-powered Crypto Market Analysis - empowering traders to cut through the noise and act decisively.

What it does

BullRun AI is a 360° crypto intelligence platform that: 1.) Tracks 20+ cryptos in real time (prices, news, volume) with 5-minute updates, using our Real-Time Stock Price Ingestion Pipeline hosted using Google Cloud Run. 2.) Predicts 5-day price movements (open, high, low, close) using hybrid LSTM-Attention models with accuracies of more than 90%. 3.) Answers complex queries via RAG-enhanced LLMs (e.g., “Why is Bitcoin dropping?” or "Which cryptocurrency had the most fluctuation today?") with citations from 1,000+ indexed articles stored in PineconeDB, with more news articles being processed hourly. 4.) Visualizes trends with interactive dashboards, TradingView widgets, and BigQuery-powered charts.

How we built it

Tech Stack: Frontend: ReactJS dashboards with dynamic visualizations using TradingView components. Data Pipelines: Google Cloud Run jobs (hourly news scraping, 5-minute price updates). Database: Google BigQuery for News + Stock data and PineConeDB for embedding news data storage. ML Engine: Bidirectional LSTM + Attention layer, trained on professional technical indicators such as RSI and Bollinger Bands with accuracies of greater than 90%. LLM Research: Pinecone for vectorized news/article storage, OpenAI for RAG workflows.

Challenges we ran into

During the development of the BullRunAI platform, one specific bug that we encountered revolved around the integration of Generative AI and RAG into our chatbot system. We faced challenges ensuring that the AI responded accurately and contextually based on real-time news and crypto listing analysis. We needed our AI LLM to have access to real-time crypto price data.

We used RAG for accessing News data, but we cannot re-vectorize and recreate embeddings for live news price data every 5 minutes. To solve this problem, we added an intermediate step on the LLM Chat API Call: We created a SQL Query to retrieve the latest prices from our Crypto Price database in BigQuery. We then passed this information in to OpenAI as context. Thus, our resulting LLM had contextual access to real-time Crypto Price data, as well as News sources that had been vectorized.

What's next for BullRunAI

We hope to convert this project into a startup one day, that can help bring the power of AI to help more people save money and save time.

Built With

• cloudrun
• flask
• gcp
• javascript
• lstm
• ml
• python
• react