How to build, manage and retain a killer team in Deep Learning
summary
Preligens runs about 60 data scientists and the same number of engineers and claims that building a production ML org requires a dedicated data pipeline, annotation squad and reproducible tooling rather than ad hoc notebooks. The team ships real detectors for satellite and street imagery (vehicles visible down to ~36 pixels), automates training, validation and logging with Docker and ML orchestration, and hits production after roughly 25 iterations to production while keeping engineers by investing in tooling and clear career ladders and a mix of juniors and experienced hires.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:51Building a deep learning team
- 5:24Hiring and skill gaps
- 10:44Team structure and roles
- 15:00Data and production challenges
- 20:22Autonomy and iterative development
- 25:01Validation and testing
- 30:18Tooling and collaboration
- 35:03Performance and scalability
- 38:57Final advice and Q&A
related talks
35mDeepL: A Product Culture that Creates Top-Notch AI Products
27mHow to make ML teams 80% more productive.
44mAI workflows and teaming - how do teams deliver AI products?
22mBuild Smarter, Don't Just Build
33mExecutive Summit Roundtable Discussion
18mSmall AI Teams with Huge Impact
21mWhy is Production ML so Hard?
30mBuilding Teams and Culture that Support ML Innovation