About


Hello, I am Sander Schulhoff

I study computer science at the University of Maryland and do Natural Language Processing Research at CLIP (PI: Jordan Boyd-Graber, Advisor: Denis Peskov). I am interested in natural language processing, reinforcement learning, and user interface creation. I plan to graduate with a CMSC major (machine learning specialization) and a minor in STEP (Science, Technology, Ethics, and Policy).


This website is not always up to date, but my Resume usually is.


Hi again,

Recent Work


Skills
Frequently Used Languages

I use Python, C, and Google Apps Script most often for work and school.

Web Dev
I have developed websites with HTML, Javascript, JQuery, CSS, and MaterializeCSS and Bootstrap. I enjoy working with APIs like Wikipedia and Google Suites.

Design
I have worked with Adobe XD, Figma, Photoshop, Adobe Illustrator to develop a number of logos, miscellaneous art, and site designs.

Machine Learning

I have read books like Deep Learning with Python, Introduction to Reinforcement Learning (Sutton and Barto), and An Introduction to Information Theory. I enjoy experimenting with RL problems and hope to eventually apply what I learn to space travel, space mining, and planetary development.

Other languages, tools, and libraries I have worked with

PHP, MySQL, Bash, Objective-C, C++, C#, Eclipse, IntelliJ IDEA, PyCharm, Visual Studio, VS Code, Atom, Juptyper Notebook(s), Google Colab, XCode, MAMP, SQLPro, PHPMyAdmin, Unity, GDrive Scripting Environment, Mongo DB, discord.js, discord.py, Swift, Clasp, NPM,... and a bunch of others :)

Work

Currently working on:

HackAPrompt

Built Summer-Fall 2023

Organized the first global scale Prompt Hacking competition (Paper), collected 600K+ adversarial prompts and published the first ontology of Prompt Hacking. Raised 40K in sponsorship from OpenAI, Preamble, Scale AI, Stability AI, HuggingFace, and 7 other companies. Accepted to EMNLP.

Hopkins Internship

Built Summer 2021

I worked at Johns Hopkins using Microsoft \Psi and building a deep learning model for event data classification. (Codebase) (Report)

Denis Peskov's Website

Built Summer 2021

I used Bootstrap and Jekyll to build this website from scratch.

SMC Data Challenge

Built Summer 2021

I worked with a team of undergrads to develop a semantic segmentation model for the "Synthetic-to-Real Domain Adaptation for Autonomous Driving" challenge. This was a good DL learning experience and our paper was accepted into the SMC Data Challenge 2021, but our results were realistically below industry standard. (Codebase) (Paper)

Diplomacy Bot Changes V2

Built Spring 2021

As an extension to my last extension of the Discord bot, I implemented a bunch of new features. See my change report here.

Advised by Denis Peskov, Jordan Boyd-Graber

Diplomacy Bot Changes

Built Fall 2020

An extension to recent Diplomacy work (ACL Paper), I worked on a Discord bot which facilitates player communication. I implemented a lie detection model to display predictions to players (model created by Denis Peskov). My change report can be seen here.

Advised by Denis Peskov, Jordan Boyd-Graber

Semi-Automated Writing System

Built Fall 2020

I was hired to write software to automate the process of writing certain content on cards with an Axidraw. I created a UI from scratch with Bootstrap, used Electron to make it into an 'app', wrote a backend with Python+pyaxidraw, and compiled the app with Pyinstaller.

Modulation Data Collection Site

Built Fall 2020

A website built for a NLP project to collect data from users answering analogy questions while being provided with Wikipedia autocomplete suggestions. It makes use of the Wikipedia autocomplete API.

Built for Denis Peskov

Dither

Built Fall 2020

Dither is a 2D puzzle game made for ENGL290, one of my UMD courses. You play as a computer processor moving around to solve logical puzzles.

TRS Website

Built Fall 2020

A pretty basic responsive website I built for TRS.

My Website

Built Fall 2020

A pretty basic responsive website I built for myself. I recently switched to Jekyll for content management instead of a Python script I wrote which produced HTML.

Simple Continuous Neural Cross-Entropy in Project Malmo

Built Summer 2020

A draft of a video describing the training process of a simple (Pytorch) deep learning agent in Microsoft's Project Malmo I created for some advisors (Christopher Kok, Ignacio Raposo) from Chronic Coder Academy over the summer after highschool After reading Intro to RL by Sutton and Barto and some of Deep RL Hands On I wanted to apply RL to a simple situation in project Malmo. I used this videomaking process to learn Blender and Manim (3b1b's animation library). Project Malmo tweeted it here.

Teacher Recommender System (TRS)

Built Spring 2019-Present

TRS is as an end-to-end service to complete the process of matching students with their teacher recommenders for college. A couple thousands lines of Google Apps Script generate forms, send emails, make matches, and generate data reports. The Alpha version of TRS was sold in 2019 and a much improved version has been published here on the Google Workspace Marketplace. The codebase makes use of scriplets, HTMLService, frontend/backend javascript/GAS, and various other GAS features. The frontend and website are built with MaterializeCSS.

1000+ lines of code

NPTG Code Base

Built Spring 2019

A backend/business logic codebase of PHP functions for my highschool 'natural plant teaching gardens' website. Using Composer and Google Sheets API, data stored in the Sheet is accessed and displayed in the site. Users can search through plants using a simple custom engine relying on Levenstein word distances. Curl search protocol is used to scrape and save images of new plants added to the database. The frontend is still under development, but you can see a functional demo here: demo.

1000+ lines of code

Heart Heist

Built Summer 2018

A simple top down shooter game I wrote in XCode. Music and art made in Gravit Designer and Garage Band. Unfortunately, it is no longer on the App Store :(. Some of the textures were generated by scripts I wrote in Java.

1000+ lines of code