MLOps community meetup #54! Last Wednesday we talked to Laszlo Sragner, Founder, Hypergolic.

// Abstract:
How my experience in quant finance and software engineering influenced how we ran ML at a London Fintech Startup. How to solve business problems with incremental ML? What’s the difference between academic and industrial ML?

// Bio:
Laszlo worked as a quant researcher at multiple investment managers and as a DS at the world’s largest mobile gaming company. As Head of Data Science at Arkera, he drove the company’s data strategy delivering solutions to Tier 1 investment banks and hedge funds. He currently runs Hypergolic (hypergolic.co.uk) an ML Consulting company helping startups and enterprises bring the maximum out of their data and ML operations.

// Takeaways
Continuous evaluation and monitoring is indistinguishable in a well setup product team.
Separation of concerns (SE, ML, DevOps, MLOps) is very important for smooth operation, low friction team coordination/communication is key.
To be able to iterate business features into models you need a modelling framework that can express these which is usually a DL package.
DS-es are well motivated to go more technical because they see the rewards of it. All well run (from DS perspective) startups in my experience do the same.

//Links
Free eBook: https://machinelearningproductmanual.com/
Lightweight MLOps Python package: https://hypergol.ml/
Blog: laszlo.substack.com

———– ✌️Connect With Us ✌️————-
Join our Slack community: https://go.mlops.community/slack
Follow us on Twitter: @mlopscommunity
Sign up for the next meetup: https://go.mlops.community/register
Catch all episodes, Feature Store, Machine Learning Monitoring and Blogs: https://mlops.community/

Connect with Demetrios on LinkedIn: https://www.linkedin.com/in/dpbrinkm/
Connect with Laszlo on LinkedIn: https://www.linkedin.com/in/laszlosragner/

Timestamps:
[00:00] Introduction to Laszlo Sranger
[02:15] Laszlo’s Background
[09:18] Being a Quant, then influenced, what you were doing with the Investment Banks?
[10:36] “In a quant fund, models never last alone. They are never ready.”
[12:24] Do you think this can be applied in different use cases or specific to what you are doing?
[12:53] “You need to justify why your process is not stationary and you need to have a continuous update process.”
[13:35] “You need to qualify your statements that can start growing into the product mindset.”
[14:41] Do you have any thoughts of a potentially highly opinionated person?
[15:00] “Todd Underwood… He was explicit about MLOps being a product.”
[15:26] “Machine Learning and Machine Learning Operations are not the same things. MLOps is a product that is allowing the use of Machine Learning.”
[16:54] Product management in Machine Learning
[17:36] “Traditional Machine Learning is waterfall-like. It’s a big problem because most of your energy is spent on data cleaning but you only get feedback at the end of the pipeline when you deploy it into production and start making money.”
[20:07] “Apart from Deep Learning, there are no tools that are capable of flexibly incorporating user or problem feedback or insight into the model.”
[24:59] You have to be at a large company or you have to have a large team?
[26:38] What are your thoughts on MLOps products helping with product management for ML? Is it an overreach or scope creep?
[27:04] “I definitely think that (MLOps Product Manager) is a role that should exist.”
[29:57] “It’s important not to feel that this is a chore, that the Data Scientist and ML Engineers throw at less fortunate members of the team, ‘From now on, you are an MLOps Engineer.’ ”
[32:00] In the messy world of startups due to the big cost of an MVP for NLP is RegEx which means to user feedbacks it’s incorporated by tweaking RegEx?
[32:17] “I have a problem solving it with RegEx, now I have two problems.”
[32:41] “The barrier in natural language processing is much lower, so you can have an MVP with top quality models out there relatively fast and then iterate around that.”
[33:04] Does the ensemble recent models more than older models? If so, what is the decay rate of weights for older models?
[33:55] “We kept always the old versions.”
[35:40] Since the iterative management model is generic enough for most ML projects, which component of it can be easily generalized and tools built for version control?
[36:17] “Easier model creation or maybe code generation with TensorFlow will be a trend and I see a gap there.”
[36:38] Topic Extraction: What type of model do you train for that task?
[52:55] Thoughts on Notebooks
[53:34] “I don’t hate notebooks. Let’s be clear about that. I put it this way, notebooks are whiteboards. You don’t want your whiteboards to be your output because it’s a sketch of your solution. You want the purest solution.”

source