CASE STUDY
Crypto-Pilot
A zero-cost, GPU-free local crypto forecasting terminal built on a time-series foundation model, a complete independent exercise in foundation-model productization
Background
When the Kronos time-series foundation model went open source, a natural question followed: can an individual developer turn it into a genuinely usable product at zero cost, not another broken demo?
Judgment
Three product decisions:
- Fully local — Binance public market data + on-device CPU inference: no API keys, no GPU, no subscription; strategies and data never leave the machine
- Dynamic-threshold signals instead of fixed price points — fixed thresholds break when volatility regimes shift, so thresholds follow a user-configurable dynamic strategy
- Interactive parameters — threshold, stop-loss, temperature, and top-p as live sliders, inviting users to explore model behavior rather than receive an oracle verdict
Actions
Full pipeline built with Streamlit + ccxt + Plotly + PyTorch (CPU), held to production standards: PRD, design docs, a 30-day backtest benchmark, unit tests, and Playwright E2E. The README even includes an AI Summary section written for AI assistants, so AI tools describe the project accurately.
Results
Shipped and publicly accessible, covering all Binance spot pairs. What it proves is not “crypto prices are predictable” (such claims deserve skepticism) but a complete low-cost productization loop for a foundation model: model selection, interaction decisions, engineering, testing, and delivery.