kylewired

About

I'm a software engineer in NYC and currently working as a founding engineer at a research lab focused on immersive media, where I'm developing a platform that reframes any footage for formats it wasn't shot for. The pipeline I built from scratch has been used to run pilots with IMAX, A24, Sphere, and Cosm. It has even been featured on The Joe Rogan Experience.

I've shipped multiple full-stack, production-ready projects end to end, from TypeScript and React front ends, Node and Python backends, GPU inference pipelines, device integrations, and the agentic system our team uses to test code, open pull requests, and review them, to C++ when needed. The work has also stretched past engineering, into forward deployed work with customers, product strategy and UI/UX design, marketing and GTM, and pitch decks and investor updates.

Before software, I was on the investment banking recruiting track, but pivoted in the middle of my junior year, self-taught. I joined the lab with no real experience and went from trial hire to intern to full-time founding engineer in three months and established myself as core to the team, finishing my degree while working on-site.

In my free time I play games competitively.

Anecdotes

Pivoting From Investment Banking

I was on the investment banking recruiting track through most of college and left it in the middle of my junior year to join a startup, with no engineering experience. A side project I was working on at the time is what led to the interview. Full story here.

Saving Thousands in Compute and Labor

Our generated video kept coming out with artifacts in areas dense with fine detail, mostly pebbles or splashing water. The team's plan was to train a LoRA to fix it, and that was no quick fix. Creating a proper dataset is difficult on its own, and on top of that came the compute costs and the hours of labor a full training run takes. I knew the pipeline end to end and didn't think it was a generation problem. The artifacts only showed up during lower resolution outputs, which pointed at the VAE, one of the first steps in the pipeline. A VAE compresses each frame into a smaller latent space and decodes it back out, and fine detail is exactly what that compression drops first when there aren't enough pixels carrying it. If the detail was already lost at the encode, nothing generated afterward could bring it back, so training a LoRA downstream would never fix it. I asked to hold the training work while I checked, and about an hour of testing confirmed it. What removed most of the artifacts was using a higher resolution source video and upscaling it when needed, so the encoder had enough pixels to keep the detail. In the end we shipped a config change instead of a training run that would have cost thousands in labor and compute.

$30k From a Clear Edge

When online gambling was being legalized state by state, sportsbooks and casino sites were competing for new users with sign-up promotions like deposit matches, bonus bets, and odds boosts. I happened to live in New Jersey, one of the few states where both were fully legal, so nearly every operator was live there with a new-user offer. It started when I scrolled past an ad for one of those bonuses. The offer looked too good to be true, so I ran the numbers, and it was essentially free money. If you read the terms carefully, many of those offers carried positive expected value (while others were bait with negative expected value), and some could be structured so the payout was locked in regardless of the outcome. The expected value on a few of them was so high it was hard to believe the sites were profitable running them. It took little capital and no special access, so anyone could have done it. The main thing was recognizing there was a clear edge at all. Each offer went into a spreadsheet with its expected value, nothing got played unless the math cleared, and it all took only a little time each day. At one point FanDuel held back $1,000 in credits, and pointing support to the exact clause in their own terms got them released. In about half a summer it added up to roughly $30k. Some sites eventually tightened their terms. Others never did, and as far as I can tell some of those offers are still available, which suggests not enough people have capitalized on them.