PhD student at UC Berkeley
I build machine learning systems for speech, sustainability, and resource-constrained computing. My work mixes research, product engineering, and a stubborn fondness for tools that help real communities.
I am a CS PhD student at UC Berkeley, advised by Professor Gopala Anumanchipalli, after a joint BS/MS in CS and Math at UW.
Robust phonetic transcription and pronunciation learning for non-native, accented, and impaired speech.
Hardware-in-the-loop benchmarks, tiny models, and multimodal agents for environmental impact.
Field-facing systems for farmers, students, musicians, and educators.
Research papers, startups, talks, and tools aimed at practical computing for language, sustainability, education, and access.
I founded Koel Labs to make pronunciation feedback and speech technology more inclusive. We build open speech models, evaluation tools, and products for non-native and impaired speech, supported by Mozilla, Microsoft, Google, and Nvidia.
A commercial version of an attendance system I made for the Seattle Youth Symphony Orchestras. I co-founded this company with Claire Li. Quickly take attendance using QR codes. Features OAuth2 integration, electronic payments and a powerful admin UI.
Co-founded a tutoring company with Claire Li. I've taught 20+ students in K-12 topics and music/violin. I also fascillitate peer-reviewed research opportunities for high schoolers.
First-author Interspeech work on robust phonetic transcription for accented and non-native speech. We study how human and grapheme-to-phoneme supervision scale, and ship practical models through Koel Labs.
Pre-print on enduring performance gaps in ASR systems between standard and minority English varieties, with collaborators across CMU, UBC, and international speech research groups.
Co-first-authored CHI paper on designing and scaling a dual-platform agricultural technology for 10,000 farmers in rural Tanzania.
Co-first-authored Nature Electronics work on multimodal AI agents for sustainability assessment of consumer electronics.
Co-first-authored benchmark and harness for iterative optimization via hardware feedback in embedded machine learning deployments.
Worked with my friend, Austin Ulrigg, to design a novel graph embedding algorithm optimized for cage graphs. The problem lies at the intersection of graph theory, combinatorics, and topology and has applications in network design, chemistry, infrastructure planning, and more. Accepted to Discrete Mathematics.
A visually rich exposition to topological graph theory, graph embedding, and rotation systems culminating in a deeper understanding of the novel Practical Algorithm for Graph Embedding (PAGE) that I co-invented.
It is ever more important for us to understand the environmental impact of the products we buy. Yet the traditional Life Cycle Assessment requires extensive domain expertise, data access, and time. This browser-based prototype uses in-context natural language and visuals to democratize understanding of environmental impact.
We presented our progress developing an MVP for pronunciation learning at Mozilla Builders '24 Demo Day following 12 weeks of intense user validation, machine learning, and HCI research.
Invited DUB Research Day talk on in-context interactive environmental impact communication.
Paul G. Allen School Annual Research Showcase talk and poster presentation on multimodal sustainability assessment.
Guest talk for the University of Washington embedded systems capstone on computer vision deployment for small devices.
Informal 10 minute talk on deploying speech technology at scale including techniques to optimize compute performance for streaming inference and practical concerns such as annotator bias.
University of Washington Research Symposium presentation on accessible, privacy-preserving concussion care.
Northwest Undergraduate Mathematics Symposium talk on graph embedding algorithms and optimization.
Presented our speech research and state of the art non-native speech phonemic transcription model at the Natural Language Processing poster session.
DubHacks Next Demo Day presentation on inclusive speech technology and Koel Labs.
Presented about the challenges of evaluating and building inclusive speech technology based on research with Samir Farhan (British Columbia) and Eunjung Yeo (Carnegie Mellon).
Presented a non-technical hands-on introduction to graph embedding at the Undergraduate Research Symposium with my co-author Austin Ulrigg.
Presented our work on Koel Labs at the DubHacks Next Demo Day. We did a poster presentation and were invited to do both a Headline Talk and a pitch to Madrona Venture Group.
Invited by Professor Bender to give a 50-minute lecture on applications of computational linguistics for pronunciation learning to the PhD students at the Tree House seminar.
Introduction to my company, Koel Labs, and our plans for pronunciation learning. Presented to the other 13 cohort teams at the Mozilla Builders Accelerator Kickoff Event in New York.
10 minute in person talk at Para.chi.dub about designing and scaling dual platform information dissemination systems in rural Tanzania based on my work on the eKichabi v2 project.
Professor Anderson invited me to give this 50 minute lecture at the CHANGE Seminar for 40 PhDs based on my research in the ICT4D space as part of the eKichabi v2 project.
As Tech Lead, I led development of customized assessment software to enhance Cascade Enrichment's K-12 tutoring curriculum in collaboration with Claire Li.
Data processing tools for the University of Washington's Design Build Fly team. Helped our team win 3rd place in the 28th national DBF competition.
My old youth orchestra (SYSO) needed a way to automate taking attendance so I made a system that allows musicians to scan their QR code at a scanning station (any authorized smartphone).
One of the challenges we faced in Science Olympiad was standardizing the way teams test the WiFi antennas they build. So I created a program with a nice GUI to make this easy.
This chrome extension enhances toebes.com to allow autogeneration of Cryptographical Puzzles (Code Busters tests for Science Olympiad to help my teammates practice).
NPM package to automatically manage SQLite database schema changes through a Django-like migration system but framework agnostic and with JavaScript+CLI integration. The library solves the issue of updating a production database schema without losing data or having to manually write/execute SQL queries. Used in production by attendqr.com.
A thin JavaScript wrapper around the ONNX Web Runtime that provides a nice API for loading and running inference with State of the Art Automatic Speech Recognition (ASR) models from Huggingface. I created this for my own projects because having a way to easily run the latest models locally in the browser is super convenient, e.g., for quick demos like this.
Latex has a super useful feature (synctex) where you can click anywhere in the rendered PDF to place your cursor at the corresponding location in the source code. HTML has a nice 'contenteditable' property that lets you edit the rendered website and you can then easily paste the results back into your source code. Chrome dev tools let you edit styles, attributes, and HTML directly on the website. What if you could do this with any React app using any framework? With the React Live Edit library you can!
A mini client-side web framework that allows creating webcomponents and managing reactive state with signals directly in HTML. I created this to address the framework shortage in Javascript. No such thing as too many frameworks to learn. Afterall, niche frameworks like this provide job security which is ever more relevant given recent AI developments.
Made as an intern at Gooey.AI, this quickly became their top 3 tool with 30K+ uses within the first two weeks and has landed branding deals and social media exposure for the company.
3D rendering engine written completely from scratch (down to the drawing of individual pixels) using only the CPU (so no help from the graphics card). This project taught me a lot about the math behind projection matrices, lighting and shadows, 3D transformations, texture mapping, and optimizing algorithms (parallel triangle drawing routines, frustrum culling, etc.).
This Science Olympiad challenge involved building and programming a sensor and its circuitry from scratch (using only basic circuit components and DIP) to measure salt concentration in water samples. I used a Raspberry Pi microcontroller to communicate via the I2C protocol to a NAU7802 ADC chip reading a voltage divider circuit. With proper callibration, probe submersion, and luck it is capable of less than 1% measurement error.
Simple, advertisement free, 100% local, and privacy friendly tierlist maker.
A classic web-based Tetris game with no ads. Can be downloaded and used offline.
A word game inspired by the popular Wordle but with ocean themed background graphics. Coded in collaboration with Claire Li. I coded a similar game on my own where you have to guess the state/country of the absurd law, or in a different mode select the AI generated law.
A space themed Shoot Em Up.
In the Nand2Tetris course, you build a 16-bit computer from nand gates and write a compiler for a custom "Jack" programming language. For my final project, I implemented a Minesweeper clone in the Jack programming language and ran it on my computer. The project features a modified implementation of the RNG algorithm used in TI calculators
In middle school, LEGO Mindstorm was my favorite toy. For one notable weekend project, I used it to create a robotic arm that I then programmed using micropython to solve my homework while mimicking my writing style using a pen. Then in High School, I used my LEGO Mindstorm skills to throw together a robot (in collaboration with Claire Li), a couple hours before a national competition. We got 2nd place thanks to the skillful control of Samarth Venkatesh!
Digital extension of the Chameleon board game where everyone but the Chameleon knows the word and have to give subtle clues so the Chameleon can be voted out without learning the word. With this extension, you can digitally receive the coordinates of the random word instead of rolling physical die. That way, the Chameleon (*cough cough* my sister), can't just memorize the coordinates.
Using an affordable NeuroPawn EEG headset for reading brainwaves and an SO-101 robot arm, my friend, Ethan Kogan, and I investigated perception representations for imitation learning. We managed to control the robotic arm directly using our brainwaves.
My friends (Eun Be, Anika Arugunta, Kevin Wu) and I developed a pipeline for analyzing videos of tennis players. It detects the serve moments, decomposes them into frames corresponding to the 8 stage tennis serve model, and screens for injury risks and potential technique improvements based on pose estimation and racket+ball tracking.
Reach me via alex@sandergi.com or scan the QR code to add me to your contacts.
