ML in Proptech - Concept to Production
Download as PPTX, PDF0 likes141 views
The document discusses the evolution of proptech since its establishment in 2018, focusing on data engineering, recommender systems, and the challenges faced in adapting to customer expectations. It highlights the importance of data quality, digitalization, and the integration of machine learning in improving user experiences despite traditional industry resistance. Key lessons emphasize that advanced technology isn't always necessary for business value and that adapting to change in a slow business environment can be difficult.
ML in Proptech - Concept to Production
- 1. 1 ML in Proptech - Concept to Production
- 2. 2 - Established 2018 - Milestones - Future Hello, who dis? - Dev - Data Eng @ scale - FP monkey Boyan Demir - XP with SL, now into RL - Armed bandit - Afraid of commit(ment)
- 3. - Every adventure starts with a leap of faith - Fun with pipelines - A look inside the data - Recommender Systems tl;dr - How to solve it - How to ship it - Results - Lessons learned The Story Today 3
- 4. 44 Real Estate Inside/Out - Old business, old tech - Riding a giant wave - Resistance to using tools …but the customer is changing - Need information abundancy - Expect being served 24/7 - It’s all about the me me me
- 5. 5 It all starts with dubious design… (SQL/Normalization) …goes through a phase of magic… (Clojure/Kafka) …into the data lake
- 6. 6 Data Exploration - Lack of culture around the importance of keeping a record - Register the absolute minimum required - Free input is the devil - Single day journey - Migration from obsolete systems without normalization
- 7. 7 Meme slide goes here / It’s time for MACHINE LEARNING* *if on presentation - it cures cancer and fights aliens
- 8. 8 How Netflix(Youtube) do(n’t) it - Needs data. Lots of it - Cold start problem - Recs based on user behaviour CF - Cold start for new users - Feature based similarity - Model per user, not a shared one CB
- 9. 9 The Recommender we need, but not the one we deserve 9 - Constant influx of new users - Very long repeat customer cycle - Perishable items - Limited interaction between user/items - Buy/Rent are very different paths - There is no such dataset(yet)
- 10. How to solve it 10
- 11. Here We Discuss The Future 11
- 12. Microlibs Pattern - Notebooks to Library ML Cores (serialized) HTTP to he max How to ship it 12
- 13. The Effect - Served ~ 17000 customers, via 4 channels(mailing, sms, chatbot, mobile) - ~35% open rate of communications with an average around 2.4 clicks per open - ~16% re-activation rate (inquiry/viewing) 13
- 14. Lessons Learned - Integrate high in the stack - Data quality >>>> Model complexity - Business value does not need the most advanced tech - Digitalization should ideally precede advanced ML - Available tools will not always fit every specific problem - It’s hard being a fast company in a slow business world 14