On Building Production-Quality AI Systems
summary
Databricks announces specific production features: zero-code fine-tuning for open-source models, a hosted vector index with state-of-the-art embeddings, a Tools catalog plus agent SDK that deploys and evaluates agents, and a central Mosaic AI Gateway for governance. FactSet moves from 50% accuracy and 15s latency to 85% accuracy and 3x faster by tuning models, adding retrieval and tool logic. Shutterstock ships a proprietary image model; mlflow 2.14 adds tracing for per-interaction debugging. Over 200,000 custom AI models are built on the platform.
AI-written summary · about
chapters (11)
- 0:00Production-grade AI systems
- 2:53General AI to data apps
- 6:45FactSet’s compound AI system
- 9:11Zero-code fine-tuning
- 10:57Pre-training on enterprise data
- 12:21Retrieval augmented generation
- 13:48Mosaic AI Tool Catalog
- 17:40Evaluating AI application quality
- 19:48Demo: cookie franchise AI agent
- 35:20Key takeaways
- 36:22Block’s use case with Mosaic AI
related talks
19mDatabricks
49mBuild & deploy generative AI apps in days with Databricks Lakehouse AI
16mOperationalizing AI in Finance: Real-Time Analytics
28mAccelerating AI at Scale with Data
31mBringing AI to Your Data
11mUsing AI to extend the range of data scientist
40mInsights and Key Learnings on Our AI Journey
49mGenAI for Business on the Modern Data Stack