Model deployment made easy with PMML
- 1. #1 New York #1 Software #7 Overall Rank #1 in 2015 Deloitte’s Technology Fast 500 #1 in 2015 CRAIN'S New York FAST 50
- 2. MODEL DEPLOYMENT MADE EASY WITH PMML
- 3. We understand how to connect data to human needs to create fulfilling mobile moments Our solutions Advertising Platform Digital ContentSocial Data Platform Virtual Reality
- 4. Data team 20% of employees are data experts Holding BSC, MSC and PHD, Intelligence units 8 Data scientists
- 5. About me Physics Ph.D Superconductor-to-insulator transition in ultrathin films Converted to data science 1.5 years at StartApp Optimization (ad -> user) Deployment
- 6. Data Real time 2.5B Data Requests served daily 200K/s model evaluations 25 milliseconds Response time 15TB of data per day
- 7. Request Enrichment Candidate ads Scoring Delivery Total 25 ms Real-time platform work
- 10. Using a developer Knows system architecture Can tailor code to scale May be a nice guy! Cons Usually won’t know models Translation of model proves to be difficult Long time to production What if something doesn’t work – who’s to fix it? Pros
- 11. HIGH MODEL UPDATE FREQUENCY Weekly, daily, hourly HIGH THROUGHPUT 200K requests/second LOW LATENCY Max. 25 ms end-to-end VARIOUS MODELS LR, RF, etc. Our challenges
- 13. Maintained continuously for 20 years Industry-oriented XML format Declarative Many models: Ensembles (!) Regressions Trees Neural networks K-nearest neighbors What’s PMML, you ask?
- 14. What’s PMML, you ask? Maintained continuously for 20 years Industry-oriented XML format Declarative Many models: Ensembles (!) Maintained continuously for 20 years Industry-oriented XML format Declarative Many models: Ensembles (!) Regressions Trees Neural networks K-nearest neighbors
- 15. What’s PMML, you ask? Regressions Trees Neural networks K-nearest neighbors Maintained continuously for 20 years Industry-oriented XML format Declarative Many models: Ensembles (!)
- 16. Real-time platform work Input Validation Pre- processing Model scoring Post- processing Output
- 17. r2pmml JPMML Transformers More than meets the eye
- 18. When PMML is done PMML PMML 1 2 43
- 19. LIVE DEMO!
- 20. Siesta The gap - closed Platform developer Data scientist Explore data Build feature set Train model Deploy model Measure
- 21. THANK YOU :) We’re hiring! www.startapp.com/careers/ shachaf.poran@startapp.com
Editor's Notes
- #4: Soda: social data platform Ad platform: ad platform Immerse: One of first companies to embed ads in VR. Digital content: Help physical brands connect to the mobile digital world.