As part of my remote learning technology project, our team needed cheap and maintainable infrastructure that could host and locally serve the educational content library we had amassed. I prototyped a process for converting any donated laptop into a lightweight Linx-Mint server, utilizing a RACHEL network layer for resource distribution.

We then battled Linux to optimize the OS for power consumption—to ensure our device could maintain charge for an entire classday. Ultimately 4 RACHEL PC's were deployed in SL for 2023 school year, and this project won the Mascaro Center for Sustainable Innovation Student Prototyping Grant.

After attending a few events in the video space, I identified that founders of subscription web-based communities were consistently annoyed with Vimeo and Wistia's outrageous video hosting/streaming costs and sharky overage fees. Thanks to the owned and operated cloud video infrastructure we had already built at Videonest, I was able to whip up a quick SDK, so any web developer could upload a large video file and get back a simple Videonest Embed, complete with adaptive bitrate streaming for their website.

As the popularity of the videonest-sdk grew, I added video re-encoding, subtitle generation through Faster Whisper, and ultimately an enterprise tier of the SDK with tokenized authentication.

Annoyed at Youtube's gate-keeping of creator content, I co-engineered a system designed for mass content download. I purchased and configured four Radxa SBC's and turned them into secure proxies (aka hooked them up to all of my friend's routers). I then stood up a server that uses the yt-dlp package for video extraction and tailscale to securely communicate with each proxy.

After realizing that Youtube quickly flagged non-authenticated requests, I had to implement session rotation of cookies from a series of Google accounts. This multi-node approach enabled high-throughput processing for downloading a creator's entire video library within a matter of hours.

This was a university-sponsored project for PPG Industries, where each team was given a noisy data sample of paint corrosion simulations and asked to build a model that could predict and thereby limit corrosion over time. I conducted EDA feature analysis on a variety of provided chemical and manufacturing inputs and then tried my hand at architecting and training a neuralnet, gradient boosted tree, and random forest model.

Eventually, I landed on the neuralnet model and then tuned for the highest AUC score, which ended up scoring ~93% on the test set.