Creating an AI-Python Infrastructure that works with a reasonable scale
summary
A small team builds an end-to-end Python AI stack in 18 months and puts it into production on-premises, serving about 50k users per week and integrating with a catalog of tools for ~15M authors. The system uses Hugging Face models (Specter), PyTorch, FastAPI backends, a Streamlit AI testing portal, Dockerized services, GitHub CI/CD and a custom data_connector + Airflow YAML pipelines approach; human evaluation and documentation are mandatory and the setup is operational and evolving toward Kubernetes.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 0:33Goal
- 1:24Reasonable Scale Company
- 2:41Data is King
- 4:24Machine Learning Trends
- 7:04AI Portal
- 8:35How they started
- 11:25Architecture
- 15:41MLOPS
- 16:44GitHub
- 17:46AI Content
- 18:06Copilot
- 18:34AI Architecture
- 19:51Conclusion
- 21:53Questions
related talks
24mSovereign Data for AI with Python
25mRaising Python quality across a large & diverse AI org
30mAI-Assisted Creative for Automated Marketing using Python
28mScaling AI Infrastructure
35mCase Study: AI-as-a-Service on Kubernetes at Scale and In Production
54mArtificial Intelligence seen from the software development lifecycle perspective
16mKey Points and Lessons Learned in AI Projects in Mechanical Engineering
17mDevOps for Private AI: A Scalable Platform Using Kubeflow