2025/12/06 –, Main Stream 言語: English
Forecasting isn’t just about numbers, it’s about making the future understandable and actionable. TimeCopilot (https://timecopilot.dev/) is an open-source Python agent that combines large language models with time-series foundation models to automate, explain, and interact with forecasts. In this talk, I’ll share how we orchestrate Python-based agent workflows with tools like PydanticAI and Hugging Face to bridge the gap between statistical accuracy and human-friendly insights. We’ll walk through real examples, lessons learned from building for the open-source community, and how agentic systems can make forecasting more accessible to everyone.
TimeCopilot is built in Python to bring together two powerful worlds: the reasoning capabilities of LLMs and the precision of time-series foundation models like AWS Chronos, Google TimesFM, Salesforce Moirai, and Nixtla TimeGPT. It’s model-agnostic, supports ensembles for more robust results, and lets users ask questions about their data in plain English.
This session will cover:
- How Python tools (PydanticAI, PyTorch, Hugging Face) enable agentic forecasting workflows.
- Designing prompts and orchestration pipelines that integrate LLM reasoning with domain-specific models.
- Making forecasts explainable and interactive for technical and non-technical audiences.
- Lessons from our open-source journey, including our 1st-place on the Salesforce GIFT-Eval benchmark.
Whether you’re into data science, AI agents, or building inclusive tools, you’ll leave with ideas and code you can use to create your own Python-powered forecasting agents.
🇲🇽🏳️⚧️ Machine Learning and AI leader with a background in econometrics and mathematics. Developer of TimeCopilot. Co-creator of the Nixtlaverse, and TimeGPT. She served as CTO and Co-Founder of Nixtla. She loves live coding and has spoken at events such as ISF, Microsoft Build, AWS reInvent, and PyCons. Author of peer-reviewed papers and recipient of prestigious awards in forecasting and AI.