About me

I am a current undergraduate at the Georgia Institute of Technology, majoring in Computer Science and Mathematics. I have extensive experience in research, where I've studied adversarial attacks on image generators as well as the effects of progressive sharpening in optimization. I also have experience in software development, where I've worked at the Parsons Corporation to use an ensemble of AI LLM agents (GPT4o and LLaMA 3.1) at the same time to optimize the performance of a system. I am currently an Machine Learning Engineer at Vytal.ai, where I am working on AI for medical gaze tracking.

I enjoy lifting weights, playing IM sports, and advocating politically in my spare time.

My current projects

  • design icon

    Parsons Corporation

    I deployed an ensemble of AI LLM agents (GPT4o and LLaMA 3.1) at the same time to optimize the performance of a system.

  • Web development icon

    Progressive Sharpening

    I'm researching the effects of progressive sharpening in optimization of AI models. More information forthcoming on details.

  • mobile app icon

    Vytal Gazetracking

    I'm developing on AI approaches to make medical gaze tracking more precise and generalizable.

  • camera icon

    3D Model Generation

    I finetuned weights from Jun et al. 2023 to make 3D model generation more efficient and accurate. I also developed novel methods to clean and alter generations.

Resume

Education

  1. University school of the arts

    2007 — 2008

    Nemo enims ipsam voluptatem, blanditiis praesentium voluptum delenit atque corrupti, quos dolores et quas molestias exceptur.

  2. New york academy of art

    2006 — 2007

    Ratione voluptatem sequi nesciunt, facere quisquams facere menda ossimus, omnis voluptas assumenda est omnis..

  3. High school of art and design

    2002 — 2004

    Duis aute irure dolor in reprehenderit in voluptate, quila voluptas mag odit aut fugit, sed consequuntur magni dolores eos.

Experience

  1. Creative director

    2015 — Present

    Nemo enim ipsam voluptatem blanditiis praesentium voluptum delenit atque corrupti, quos dolores et qvuas molestias exceptur.

  2. Art director

    2013 — 2015

    Nemo enims ipsam voluptatem, blanditiis praesentium voluptum delenit atque corrupti, quos dolores et quas molestias exceptur.

  3. Web designer

    2010 — 2013

    Nemo enims ipsam voluptatem, blanditiis praesentium voluptum delenit atque corrupti, quos dolores et quas molestias exceptur.

My skills

  • Web design
    80%
  • Graphic design
    70%
  • Branding
    90%
  • WordPress
    50%

Portfolio

CloudGen

Actor Critic LLM System


We (I along with two friends) developed a 3D CAD model generator designed for 3D printing. It used a diffusion model operating on NeRFs (neural radiance fields) to generate 3D models from text or images.

Here are some samples:

gif of object gif of object gif of object gif of object

These were all generated from text prompts, such as "hammer" or "jail bars"


We also designed a novel method to clean up the generated models, which were often noisy and not very accurate. We used a bilateral filter to "clean" the models, and we dynamically reversed the diffusion process to edit the models with new prompts.

Here is an example of the editing process: Starting with prompt "cow", we get a typical cow model. Then, we reverse the diffusion process, slightly destroying the model, and then, with prompt "With horms", we get a new cow with larger horns.

cows
  • A NeRF: a function that represents a 3D model. Think of it like f(x, y, z) = (color, density), but it is a bit more complex than that.
  • Diffusion: an AI training process that starts with a dataset, "destroys it" by adding noise, and then trains a NN to recover the original training data, effectively making an AI that can turn noise into synthetic pieces

I designed an Actor-Critic LLM System to respond to user's questions on ParsonsGPT, where I interned Summer 2024.


While LLMs like GPT4 (ChatGPT) have shown incredible success, they still suffer from issues in hallucinations and inaccuracies, along with often producing general sounding outputs. However, where one model fails, another model may succeed, and so a growing amount of literature is showing that when multiple models are used in conjunct, they perform better


Say you wanted a description of the gradient, and you ask GPT4:

gpt4 first time

Now, this works, but it is brief, and you want more. You ask a second agent, LLaMA 3.1 to critique the answer

llama first time

Now, LLaMA has issued a critique, and GPT will incorporate the feedback.

gpt4 second time

Now, GPT has issued a more detailed response, and if you wanted, you could ask LLaMA to critique it again.

Blog

Contact

Contact Form