Today's thread - by @pingali - is about making #MachineLearning projects successful. Links back to the first session of making #DataScience work by @scribbledata hasgeek.com/fifthelephant/… (1/8) 
@pingali @scribbledata Framing: Formulating a good problem is the most important step. What is 'good' depends on the context. Mature companies require optimal solutions. Sometimes, framing can be good enough but something that can be pulled back. Take #risks, with guard rails in place. (2/8)
@pingali @scribbledata Buy-in: #DataScience has ambiguity. Communicating is
critical - define goals of communication; who you are communicating with is important. Link data science outcomes to business outcomes. Understand incentives and motives of other stakeholders. Be empathetic. (3/8)
critical - define goals of communication; who you are communicating with is important. Link data science outcomes to business outcomes. Understand incentives and motives of other stakeholders. Be empathetic. (3/8)
@pingali @scribbledata #Team: Evolve it according to the stage. Roughly 4 roles: #product,
#data, #engineering, and #modeling. Very early startups could have one hands-on person doing all. Mature will have two for each role, with at least one experienced person. (4/8)
#data, #engineering, and #modeling. Very early startups could have one hands-on person doing all. Mature will have two for each role, with at least one experienced person. (4/8)
@pingali @scribbledata Ramp up: In early stages, ramp up takes place when data engineering comes before #DataScience. Convert raw #data into usable, structured data stores. For larger organizations, start with central team, and build function-specific teams over time. (5/8)
@pingali @scribbledata System design: Build ML infra that can help with solving multiple problems, but with guard rails. You have to keep options open because you don't know which one will 'catch fire'. #Automate early. Standardize. Embed lessons into code. Makes #experimentation
more efficient. (6/8)
more efficient. (6/8)
@pingali @scribbledata #Operations: Bad things happen in production. End-to-end #infrastructure has to be aligned to #DataScience process. Engagement with #engineering starts early to make sure right #data is collected. Build measurement, circuit-breakers, fallbacks, guard rails around models. (7/8)
@pingali @scribbledata Takeaways: (a) Engage stakeholders and have seat at the table (b) Align for fast experimentation and learning (c) build for #automation, standardization and asset reuse. Read summary and watch video on: hasgeek.com/fifthelephant/… (8/8)
