Build a personalized Bitcoin (BTC) virtual assistant in Python with Hopsworks and LLM function call…

Built a Bitcoin virtual assistant using Hopsworks, Python, and LLM function calling that can analyze trends and make price predictions

System architecture follows FDI Pipeline with three components:

• Feature pipeline: processes Bitcoin price data from Binance and Twitter APIs
• Training pipeline: trains TensorFlow model for price predictions
• Inference pipeline: handles model deployment and real-time predictions

Uses function calling to allow LLM to:

• Retrieve historical price data
• Make price predictions
• Process user queries intelligently

Implements Retrieval Augmented Generation (RAG) to provide LLM with contextual data and real-time information beyond its training cutoff date

Uses OpenHermes 2.5 (Mistral-based) as the LLM, fine-tuned with ChatML format

Leverages Hopsworks for:

• Feature store management
• Model registry
• Model serving capabilities
• Feature group organization

Function calling proved reliable when properly structured with clear examples and instructions

System limitations include:

• LLM can sometimes get confused with previous questions
• Price predictions should not be considered financial advice
• Market volatility makes predictions unreliable

Uses prompt engineering and templates to guide LLM responses and function selection

Integrates streaming pipeline for continuous data processing from Binance and Twitter