Skip to content

crewai-stock-analysis

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

crewai-stock-analysis

README.md

CrewAI Stock Analysis

Multi-agent AI stock analysis powered by local LLMs on remote GPUs.

Why GPU CLI?

  • $0 LLM costs: Run 32B parameter models locally vs $0.03/1K tokens for GPT-4
  • Privacy: All data stays on your GPU - no API calls to OpenAI/Anthropic
  • Fast: GPU-accelerated inference is 5-10x faster than CPU
  • Zero setup: No need to install Ollama, download models, or configure anything

Quick Start

gpu use .

This single command:

  1. Provisions a GPU pod with an NVIDIA RTX 4090/A40/A6000 (24-48GB)
  2. Installs Ollama and downloads the appropriate model (first run only)
  3. Launches the web API server for stock analysis

Once ready, analyze stocks via the API:

# Analyze a stock (streams progress in real-time)
curl http://localhost:8501/analyze/NVDA

# List completed reports
curl http://localhost:8501/reports

# View a specific report
curl http://localhost:8501/reports/NVDA

Single query mode - analyze one stock via CLI:

gpu run python main.py NVDA

What Happens

During Setup (automatic on first start)

  1. GPU CLI provisions a pod with an NVIDIA RTX 4090/A40/A6000 (24-48GB)
  2. Ollama installs automatically during provisioning
  3. The appropriate model downloads based on available VRAM:
    • 48GB+ VRAM: Qwen 2.5 32B (excellent reasoning)
    • 24GB VRAM: Qwen 2.5 14B (good balance of speed/quality)
  4. Configuration is saved for subsequent runs

During Analysis

  1. Three AI agents collaborate to analyze the stock:
    • Research Analyst: Gathers news, market data, and company info
    • Financial Analyst: Analyzes valuation, growth trends, and financials
    • Investment Advisor: Synthesizes everything into a recommendation
  2. The final report syncs back to your local reports/ folder

Endpoints

Endpoint Method Description
/health GET Health check (used by readiness hook)
/analyze/<ticker> GET Run analysis (streaming text response)
/reports GET List completed reports
/reports/<ticker> GET Get a specific report

Output

After running, you'll find a comprehensive investment report at:

reports/NVDA_analysis.md

The report includes:

  • Recent news and developments
  • Financial metrics and valuation analysis
  • Peer comparison
  • Clear BUY/HOLD/SELL recommendation with target price
  • Key risks to monitor

Files

File Description
gpu.jsonc GPU CLI configuration (GPU types, ports, readiness hook, environment)
startup.sh Startup script - starts Ollama, downloads model, launches web server
web_server.py Flask REST API wrapping CrewAI analysis
init_ollama.py Idempotent Ollama setup - verifies server and downloads model
main.py CLI mode - analyze a single stock via gpu run
crew.py CrewAI crew definition with agents and tasks
ollama_utils.py Ollama server management utilities
config/agents.yaml Agent role definitions
config/tasks.yaml Task descriptions and workflow
tools/search.py Web search and scraping tools

Customization

Change the LLM Model

Edit ollama_utils.py to use different models:

def select_model(vram_gb: int) -> str:
    if vram_gb >= 40:
        return "llama3.1:70b"  # Larger model
    return "llama3.1:8b"  # Smaller model

Modify Agent Behavior

Edit config/agents.yaml to change agent roles, goals, or backstories.

Customize Analysis Tasks

Edit config/tasks.yaml to modify what each agent focuses on.

Add More Tools

Add new tools in tools/search.py - for example, you could add:

  • RSS feed reader for financial news
  • Stock price API integration
  • Document analysis tools

Performance

GPU Model Analysis Time
A40 (48GB) Qwen 2.5 32B ~10-15 min
RTX 4090 (24GB) Qwen 2.5 14B ~5-10 min

First run takes longer due to model download (~10-20GB).

Troubleshooting

"Model not configured" from health endpoint

The Ollama server is still initializing and downloading the model. Wait for the readiness hook to pass - you'll see "Service Ready" in the terminal.

"Ollama server failed to start"

The Ollama installation may have failed. Stop the pod and start fresh:

gpu stop
gpu use .

"Model download failed"

Large models can fail to download on slow connections. Try a smaller model by editing ollama_utils.py.

"No search results"

DuckDuckGo may rate-limit requests. The analysis will continue with available data.

Requirements

  • GPU CLI installed and configured (gpu auth login)
  • RunPod account with API key
  • No additional API keys required!

Cost Estimate

  • Pod cost: ~$0.50-1.00 per analysis (15 min on A40)
  • LLM inference: $0 (runs locally)
  • Compare to: $5-10+ for equivalent GPT-4 API usage