likely llc

The pandemic started. China paid me a good salary, but I left China, since it was getting locked down. I sat in Bangkok, Thailand, watching the death toll go up, while practicing writing Stan models. I had a few hundred dollars from my stipend. I had to make a move. I flew back to Detroit. I couldn’t withdrawal money, since Bank of China wouldn’t let me withdrawal outside of China.

I’m in a group home in Detroit, with some other vagabonds, per usual. UESTC cut my funding, since I had left the country due to the pandemic. The local grocery store wasn’t getting regular deliveries, the food was spoiled. I was getting robbed at gunpoint since money was tight for everyone. I computed that I would have about $500 bucks after a plane flight. I just finished my second semester. I figured, it’s time to learn Spanish, given that I am I Spanish-American.

I skated to Cancun, but we’ll skip this. What happens in Cancun stays in Cancun. I got some furniture and rented a house illegally, under the table, after working with some locals. I find Mexico in general to be more friendly. When I bought a table, the shop owner was laughing. He said, in Spanish, “I was illegal in your country, now you’re illegal in my country.” There wasn’t much there, so I figured there was more opportunities in CMDX, the largest city in North America. I started taking freelance jobs. I could make my money last a month or two, and I figured I’d take the risk, and hired a small business to take my table and bed to CMDX. Mexico has a great culture of small businesses. I was in a hostel, as usual. I found a landlord, who was a doctor that lived in Cuernavaca, that was willing to rent me the apartment. I was lucky to have run into Martin Szyzclian, another developer, but more of a user-interface focus to data vizualization, that helped me get my freelance work rolling. Argentinian guy, I don’t remember how we met, at all.

Luckily, there’s a lot of small businesses that provide high quality service that’s affordable, and people are willing to negotiate and ensure you have a job and housing. In Mexico, this is cultural. USA is much the opposite.

I’m in Mexico City, I just moved into a new apartment, and I had a bed and table which I used as a desk, and my computer and sax. The move cost me money, the furniture cost me money. I had a few hundred, enough to last a month or so. I ended up getting an interview with the Rays, and I threw this together in week: GP models in Stan take a while to run. I got lucky.

Multilevel Gaussian Processes to Model Hit Speed in Baseball

This is all in Spanish.

My recent contracting boss cut me off. I signed and NDA, and I was fitting dose-response models using GP’s in Stan, and I asked for credit for the work, but he ghosted me and refused to pay. I talked to one of my former colleagues who was subcontracting from him, she was a Brazilian biologist, and she said she got paid. I eventually got him to pay out.

The GP model was simple: a logistic likelihood with a GP prior on , with multiple inputs, used to determine log odds probability of death and interaction affects of illicit substances. The model was informative. The model below:

, a Gaussian process prior, with a squared exponential covariance function on a logistic regression model: . It was informative, and accurately measured log-odds probability of death, and capturing interaction affects of combining illicit substances.

The model should be here, somewhere: https://likelyllc.com/2023/07/14/gaussian-process-model-dump-aalto-university-internship-summer-2018-part-2/

The Stan documentation can be found here: https://mc-stan.org/docs/stan-users-guide/regression.html#logistic-probit-regression.section.

But I landed an interview with Will Cousins, PhD from the Tampa Bay Rays, through the Stan Development Team network. They gave me a programming test. Although it may not be something I’d put out as an engineering work, due to some HMC sampling issues (the model I specified had issues), it showcased my skills and I produced a good demo, seen here:

Multilevel Gaussian Processes to Model Hit Speed in Baseball

The data was public, so I can share it.

Playing Money Ball

The offer they gave me was low. I had may be $800-$1200 in my account, which gave me a month or so, to survive, with some room for entertainment. That’s a lot of money for many Chilangos (term for Mexicans in CDMX). But I needed cash. I knew I was worth more than that. Baseball is a multimillion dollar industry, so I knew they were just trying to pay as little as they can and get the best value, since I was willing to take anything I could get to keep rowing the boat. I ended up negotiating about 40% higher.

This was when Arozarena was on the Rays, in 2020. I think this was the first time they advanced to the world series, since 2008. Although my models wouldn’t be in affect until next year, I was happy to be apart of a winning team.

I ended up interfacing with Taylor Smith, a great statistician, who I had actually met at the Big Data Summer Institute at the University of Michigan SPH in 2016, and we connected years later. He’s a great statistician and now the assistant general manager of the Red Socks. They didn’t keep me because, “…you don’t know anything about baseball.” I’m a modeler and developer, so the final product came out well. My preferred sports are hitting each other in the face and choking each other out. Although, once in a meeting with Dr. Murthy, he said, “…you should stop boxing. That’s medical advice.” I boxed in CMDX a bit. They’re great boxers. Anyway…

I left CDMX, for some reason, my instinct told me it was time to go. I ended up in a studio in Southwest Detroit. It was rodent infested. I set up mouse traps. I killed 11 mice. I counted them. They got progressively smaller. First, huge eared mouse with a wart. Then more. When we got to the final ones, I caught two babies in the same mouse trap. I’m guessing their providers died, and they had to find food themselves. But anyway.

I ran so many Bayesian models in Stan that winter (I think 2020/2021), I melted the motherboard on my computer, and I had to get it replaced. The final model predicted really well, but it’s a matter of integrating Bayesian models with Stan into a corporate style work-flow, so I hope they were able to use it. It worked but was a little practically infeasible due to required computational skills and domain knowledge. But Taylor Smith is good, and he’s now the Assistant GM of the Red Socks, a well deserved position. Good job, homie.

Aside, I listen to sports radio. I’m an athlete but more concerned with my career, but while in Detroit, I was listening to the MLB draft jargon. Since they had one of the lowest budgets in baseball, they’re gonna lowball you. But I needed the money. One of the guys said,

“If the Rays call you, you hang up the fucking phone.”