| I've had that title. It helped that I got a PhD in Physics so I knew a lot of math already, also I had been interested in ML for text classification circa 2005 or so. From 2001 to 2010 or so I was busy with a wide range of web applications such as: a social network for a secret society, a blog for a local political party with an integrated telephone response system, an application tracking system for a nanotechnology internship program, etc. There was plenty of brownfield work in there. I had worked on a series of side projects that got attention and landed me a job as a "relevance architect" (recommender systems and such for a new social media site that softlaunched but didn't get big) and then an ML software engineer where I completed a search engine for patents based on a neural network. After that I went through a phase of doing random consulting projects but also trying to start my own startup around data engineering problems that didn't get the support I needed. I learned Python because there was huge amounts of Python work in this space. I did a project involving LSTM networks for text around this time. When I threw in the towel I joined up with a company where I was between the engineering and data science teams, over the course of a year I had figured out most of why our Python systems were not entirely reliable but the company was pivoting a lot and towards the end I was writing more typescript + Scala. Even though I had the physics education I felt my long experience as an applications developer was more important to this role: we had data scientists who were better at creating models than me, but most of them worked in Jupyter Notebooks and didn't have a clear idea of how to take the code they wrote and make it reusable: how to make the "monthly sales report" as opposed to the "April sales report". It was my role to get that discipline in terms of collaboration w/ the data sci's, the process we used, and as embodied in the software I developed. It was a great experience but also a bit disorganized as we were trying to develop a product but also having to change tack every week to accommodate project work for A-list companies we had as customers. We had a system called "Themis" that was used to build training sets that worked but I disagreed with architecturally (#1 requirement is not sweating it when your A-list company needs something really different). I wrote up a description of a system called "Nemesis" which the company didn't go with but I developed a number of (image|document|comment|job advertisement)-sorters on my own account afterwards that were all called "Nemesis" until I had a vision which lead to YOShInOn and the newer FraXinus which is meant to be an everything sorter. We were working on CNN-text models at this time, BERT came out and we thought it was a big advance but we had no idea how big it would be. I was burned out from working remote so I went looking for an local job which means doing more ordinary stuff like React programming but I sometimes I get to do some more systems oriented stuff such as writing parsers and code generators and such. At the end of my ML phase I felt that (1) it was more important to have the appropriate labeled data rather than the best models, and (2) UI was the bottleneck for (1) so it was worth "getting gud" at UI so I have. I still have my side projects. |