Samuel B. Mayle

Jan 2023 — Sept 2024 · Winchester, MA

Built a graphics card on solderless breadboards with my friend Iosif, based on Ben Eater’s schematics for his “world’s worst video card.”

The design draws a picture on a VGA monitor with no processor at all. A 10 MHz oscillator drives chains of 74LS161 counters that count pixels across each line and lines down the screen. Logic gates watch for specific counts to produce the horizontal and vertical sync pulses a monitor needs, and to blank the screen between lines and frames. The same counters double as the address into an EEPROM that stores the image, and resistor networks turn each byte’s color bits into the analog red, green, and blue voltages VGA expects.

We reworked parts of how it worked and expanded on his design along the way, finishing the breadboard by September 2023 after a lot of trial and error and careful wire organization. We wanted to go on to breadboard a full computer, but stopped when high school started.

A machine-learning tutor for leining, chanting Torah with its cantillation marks (trope). It tracks a reader’s pitch, compares it with an expert’s using dynamic time warping, and gives feedback phrase by phrase.

Underneath, Trope2Vec learns embeddings of whole trope phrases from Chabad.org recordings aligned to Sefaria’s text; the best model puts the right trope set in its top five 67% of the time, against 31% by chance.

Kabbalistic manuscript research with my friend Leone, under Yitz Landes at the Jewish Theological Seminary.

We turned JTS’s microfilm guide to 1,009 Kabbalah manuscripts into a research-ready dataset. The export had lost every Hebrew letter, so we rebuilt each record from JTS’s full catalog, dated 993 manuscripts by century, matched 989 to the National Library of Israel’s catalog, and translated its Hebrew into English with a 229-term glossary. Where the two catalogs disagree, on 408 manuscripts, both readings are kept, and we proposed seven groupings of related manuscripts.

Next, I’ll be fine-tuning a custom language model to help read and catalog these manuscripts.

Research assistant to Raffaella Sadun at Harvard Business School, working on reskilling: how training can help adults move into better jobs as AI changes work.

The work is part of HBS’s new Project on Israel, which brings together researchers from Harvard and Israeli institutions. The first task is mapping the training programs that governments and philanthropies fund: who runs them, how many people they have trained, and whether they have been evaluated against a counterfactual.

Wrote a script that converts a bunch of old Windows PCs at CRLS to Ubuntu and joins them into a Slurm cluster.

With every machine on the same scheduler, small data-parallel jobs can be split across the whole cluster instead of running on one computer.

A rebuild of Aspine, the student-made grade viewer at CRLS, after new login rules stopped the original from working.

Students sign in with their school Google account, and Pine pulls grades through the district’s OneRoster API or, without district access, a one-time Aspen connection with the password stored encrypted. A browser extension links your Aspen session, and Telegram alerts you when a grade changes.

Online Euchre for when you can’t find four players, built with my friend Matt Sokoloff.

Friends play together in the browser, and empty seats go to bots that pick cards with information-set Monte Carlo tree search, a standard method for games with hidden cards, so the bots never peek at anyone’s hand. It runs on a Python game server with a React client styled after Chess.com.

Choosing classes at CRLS mostly meant asking around, so this site lets students rate and review the classes they’ve taken.

It has an Express API and a PostgreSQL database for accounts and reviews, token-based login with email verification, and runs on Vercel.

Full classes open up when someone drops, so this bot checks Aspen every minute and requests the course as soon as a seat frees up.

It replays Aspen’s own web requests and falls back to driving a real browser with Puppeteer. It asks before submitting, and a React dashboard shows whether each class is free, full, restricted, or newly opened.

An open arena for benchmarks: anyone can post a problem, and people and AI agents compete by submitting solutions, never self-reported scores.

Instead of a central judge, a committee of peers re-runs each submission in a sandbox, anyone can challenge a result, and reputation makes fake accounts expensive.

Won the Beat The Clock Agent Hack.

Drop in an Instagram reel, a room photo, or a screenshot; the app pulls frames from the video, a Subconscious agent identifies each piece of furniture and decor, and you get matching Wayfair searches. A trace panel shows how the agent reached each result.

Schematics are written for people who already know electronics; beginners need step-by-step instructions.

BreadboardGPT compiles a circuit (from a built-in library, a Wokwi diagram, or a KiCad netlist) into a parts list, a realistic breadboard layout, and LEGO-style “place this, then wire that” steps, then checks that the finished layout still matches the schematic.

Harvard photographed the whole sky on glass plates from about 1880 to 1990, long enough to catch stars that dim only once every few decades.

This is a blind search of those plates for eclipsing binaries with periods of 2 to 70 years, with each candidate’s next eclipses preregistered before any modern data is opened. So far it has found a luminous giant star that dimmed twice on the plates, for reasons still unknown.

#1 on several of NVIDIA’s SOL-ExecBench leaderboards, which score GPU kernels by how close they come to the B200’s theoretical speed limit.

The work meant hand-writing kernels in CUDA, Triton, and cuTile and running every candidate through NVIDIA’s unmodified evaluator on cloud B200s.

Gaussian splatting reconstructs 3D driving scenes from camera footage. This project uses SAM segmentation masks to give people and nearby objects priority during training.

On KITTI crops of people near cars, a priority-weighted model improved held-out PSNR on people by 0.78 dB over the same model trained uniformly.

Work in Lean 4 on an open problem from Choi, Erdős, and Szemerédi (1975): how many more than n elements a subset of {1, …, 2n} needs before it must contain all the pairwise sums of some k distinct integers.

The work focuses on k = 5, where the open question is whether that number is always 4 once n ≥ 15.

Sept 2023 — Sept 2025 · Cambridge, MA

Helped improve CRLS’s grade-viewing site since 2019; new login guidelines later limited how it could run.

June 2023 — Aug 2023 · Jackson, WY

Electric bike from MIT Swapfest and flea-market parts for spread-out Wyoming travel with less environmental impact.