Sasha Madden Ebersole

Software engineer in Seattle. I build neural networks from scratch — my Manchu OCR model is a NumPy transformer with hand-written backprop — and I'm moving into interpretability research.

## Highlight

Dots and Circles — Manchu OCR

92.9% exact-word accuracy on unseen print words (ManchuOCR: 69.2%). With no handwritten words in the training data, it scores 85.5% on a handwritten set (LLaMA-3.2-11B: 87.4%). Runs ~90 words/s on one CPU core.

Try it →Source →

## Repos

live from GitHub

repo	description	commits	pushed
dots-and-circles	Dots and Circles is an app for optical character recognition of Manchu-language text, using a transformer architecture coded from scratch with NumPy matrices.	371	—
Raschka-LLM	Working through "Build a Large Language Model (From Scratch)" by Sebastian Raschka	24	—

## Next

Completing BlueDot Impact courses and working through the ARENA curriculum.

## Reading group

all slides →

Since August 2026, a few friends and I have met to present ML and CS papers every few weeks. Every deck is in reading-group.

date	paper	presenter	slides

newest first · presenter = GitHub profile name of who added the slides

## Experience

2026–: AI Trainer, DataAnnotation — evaluating coding agents
2022–24: Partner & Frontend Developer, DataMixMaster
2023: Full-Stack Developer, Groenewold IT Solutions
2022: B.A. Computer Science & History, Hunter College