Scott Susanto

    LinkedInGitHubHuggingfaceX

    Just another engineer building AI agents good enough to replace myself.

    Deeply care about the intersection between AI engineering and ML research. Set on building browser agents that can navigate the AWS website, code agents that can rewrite my entire codebase in rust, and models that escape Mt Moon.

    If there's one thing I wanna do without AI, however ironic, it's writing - come dive deep into my world. Or, stay, and explore some nerd stuff.

    Multimodal AI

    A coding agent, browser agent, conversational audio agent - collecting my AI infinity stones. In all seriousness, spent a ton of time building these, check them out

    Surf

    CUA as a browser extension, available on the Chrome store

    🔥 flamethrower

    Your shell, with AI superpowers - 18 months before Claude Code and OpenAI Codex

    This last one is pretty cool; you can try it out right now 👇

    Machine Learning

    AI engineering is cool, but I'm deeply passionate about the research that goes behind it. Ilya Sutskever, Geoffrey Hinton, Alex Krizhevsky: it's truly their shoulders the modern world is standing atop of.

    Perpetual

    Can instruct-tuned models learn new things? In this work we explore a novel technique inspired by human ways of learning new facts, utilizing both raw information and flashcard-style questions, attempting to teach instruct-tuned models new information without losing their conversational behavior. We observe that Mamba-2.8b can in fact learn new factual knowledge while still retaining assistant behavior, confirming our initial hypothesis that instruct-tuned models can indeed continue to learn

    MedTranslate

    This work explores the simplification of medical texts to improve health literacy, especially in under-resourced regions. The study uses the MedEasi corpus and the ctrlSIM model, employing a T5-Large model for simplification. To address computational limitations in third-world healthcare settings, a novel knowledge distillation approach is used, with a T5-Small model as a student model emulating the T5-Large teacher model, with performance of the student model evaluated using various metrics, including SARI, ROUGE scores, and readability tests. While conventional metrics show satisfactory results, human evaluations reveal that the student model sometimes fails to simplify complex medical jargon.

    TrashCNN

    Recycling is a well known solution to saving landfill space, however, the sorting of trash is certainly a nontrivial process. Sorting recyclables before reaching the recycling facility is crucial for effective recycling as it keeps costs down by preventing clogged machinery and the need of manual sorting in the facilities. If contaminants were to pass, the final product would be deemed unsatisfactory and thrown into the landfill rather than being reused. This experiment's purpose is to help improve models designed to classify six different forms of waste: glass, cardboard, metal, paper, plastic and trash.

    Misc

    Hackathon projects, Lava demos, and templates you might find useful

    Experiences

    alt

    Amazon Prime Video

    Building a SOTA Data Lakehouse

    alt

    Hatchet

    Firefighting analytics

    alt

    xAI

    Grokking Grok

    alt

    Atlas

    Product Engineering

    alt

    Cofactory

    AI Engineering

    alt

    Apriora (YC W24)

    Infra Engineering

    alt

    Amazon Web Services

    Systems Engineering

    alt

    Sourcegraph

    Backend Engineering

    alt

    University of Southern California

    Computer Engineering & Computer Science

    Insert cool quote here

    LinkedInGitHubHuggingfaceX