Building Scalable ML Systems: Strategies and Insights
summary
Freshworks builds production ML for Freshworks CRM that solves cold start by stacking models — account-specific → cluster/industry → global fallback — and makes first‑purchase recommendations by chaining two graph conv nets (item→item embeddings fed into user→item model) plus matrix factorization so new users get relevant items from day one. Conversational automation uses Freddy with staged intent handlers and an internal Freddy prompt tuner to cut LLM token use versus LangChain. Models deploy via SageMaker/S3, autoscale across hundreds of thousands of tenant models, and surface interpretable lead‑score factors to let customers adjust predictions.
AI-written summary · about
chapters (6)
- 0:00Introduction
- 1:06Cold start problem
- 4:23Recommendation system challenges
- 9:43Generative AI systems
- 15:29Engineering challenges
- 19:54Q&A
related talks
23mML Platform Challenges & Applications at Freshworks
47mBuilding ML practices to address the top four use cases
38mFantastic ML apps and how to build them
29mMachine Learning in Production: What to do after you’ve trained the models?
38mReal Time ML Pipelines in Multi-Tenant Environments
27mScaling ML to the next level: Hosting thousands of models on SageMaker
32mAutomating Production Production Level Machine Learning on AWS
32mMaking Models Work: How an optimized ML lifecycle can drive better ROI