San Francisco, CA
<aside> 💡 ABOUT Interested in design engineering and programming stack: python, typescript, next+react
</aside>
<aside> 💻 SOCIALS Email: [email protected] Twitter: https://x.com/ahiajsbwks Github: https://github.com/kingsotn
</aside>
<aside> 🏫 WORK @twelvelabs CS @NYU (Grad: 2024) Computer Vision Research @Ai4ce Lab (2022-2024) SWE Intern @IBM (2023) SWE Intern @Cagetest (2021)
</aside>
https://x.com/ahiajsbwks/status/1866983483702431977
<aside> <img src="/icons/triangle-alternate_gray.svg" alt="/icons/triangle-alternate_gray.svg" width="40px" />
working on something cool
🔍 🌐 📹
…stay tuned
</aside>
<aside> 📹
**pages**
gpt typescript nextJS
I built an app that generates a blog from a YouTube video using groq and twelvelabs api
This was inspired by perplexity pages. I used NextJS pages router, but will be using app router moving forward.
I wrote some technical documentations and went on a deep dive README on the technical decisions and tradeoffs and design engineering decisions. I put a lot of time into this, and if you want to learn a little bit about my challenges with UI/UX and api calls you can view that you can view here:
https://github.com/kingsotn/pages/blob/main/README.md
</aside>
<aside>
🎙️ **voice2code (VSCode Extension)**
gpt javascript
I built a voice-to-code voice extension with Rob using gpt-4o and whisper-v3.
Unlike other extensions or code editors that use voice and AI to code, voice2code allows you to code while speaking. It does this by changing only the parts that need to be replaced in the file with regex.
⌘ + . on your keyboard to activate the extension. We made it hands free, meaning that it would know when to process your request by keeping track of the average amplitude in the sound of your voice.
https://github.com/shinpads/voice2code
</aside>
<aside>
🪨 compute-exchange.vercel.app
****frontend typescript+tailwind nextJS
Briefly worked on a DEX proof of concept that allowed users to trade their compute units. I obsessed over the user experience and small details. Taught me a lot about frontend design, particularly on contrast and hierarchical component layering.
Read my unfiltered blog here 📔
https://github.com/kingsotn/orderbook
Screen Recording 2024-06-23 at 11.00.35 AM.mov
</aside>
<aside>
🤓 **nobullshitlinkedin.com**
gpt full-stack cloud
I built an extension to summarize long-winded LinkedIn posts.
I was scrolling twitter one day and stumbled across Soren Iverson’s joke tweet for an AI LinkedIn summarizer. I was already working on an extension project, and thought that I could probably build his idea in about a day. I reached out to him with a working demo the next day, and we decided to ship the product together on producthunt.
It was a quick and fun end-to-end 2-week project in learning how to creating the app, deploying to cloud servers, creating the landing page, shipping to beta users, and marketing on twitter.
how it started how it ended
</aside>
<aside>
🖼️ **InstaLoRA.xyz**
ai/ml full-stack cloud
InstaLoRA uses your face in diffusion image generation using short videos of your face.
Both training and inference was done with Modal’s H100 serverless platform. This was a project done in collaboration with Leo Li for the ivyhacks.ai hackathon. We won first place and best front-end.
Funny story — Leo and I were never actually accepted into the hackathon, and when we got in we spent the entire first day at the hackathon not knowing what to build, and ended up actually building the entire project in about 12 hours right up until demo time.
InstaLoRA - Instant LoRA Generator
</aside>
<aside>
🤖 Spatial Reasoning Research
ai/ml research
I curated a synthetic dataset from simulation software, used in ai spatial understanding or sparse 3d reconstruction.
How do we reason that two images are of the same room? It’s a bit more complicated than you think.
This problem requires more than just feature recognition — understanding the broader context and spatial relationships of the agent and environment is crucial.
This could improve technologies like structure-from-motion (SfM), simultaneous localization and mapping (SLAM), and robot navigation.
I primarily worked on creating the synthetic datasets from simulation software, and spent a lot of time on most of the visualizations.
[https://drive.google.com/file/d/1yux5Jo09r-ONS7qLpLCANl4DZGMYb3hH/view?pli=1](https://drive.google.com/file/d/1yux5Jo09r-ONS7qLpLCANl4DZGMYb3hH/view?pli=1)
</aside>
<aside>
📐 BERT-LSH & AlgoML Grad-level class
ai/ml research
I implemented a BERT model using Locality-sensitive Hashing in the attention layer, resulting in lower loss and compute.
For the class, GPTs were just released on chatgpt, and so I ended up scraping all our TA notes, converting them to text files and creating a no-code RAG implementation of TealGPT (named after Teal, our TA).
This was the most difficult class I’ve taken at NYU, and I highly suggest other students interested in ML research or just understanding AI/ML twitter to take this class. It’s hard, but very rewarding.
https://arxiv.org/pdf/2404.08836
BERT-LSH: Reducing Absolute Compute For Attention
https://github.com/leo4life2/algoml-final
</aside>
<aside>
📊 Terminal Chain of Thought App
gpt
What if natural language was all you needed to power a backend?
That was the inspiration behind the terminal Chain of Thought app, which chained gpt calls behind functions in your terminal. Each prompt invokes another chain-of-thought, which then invokes another function, or another thought until the task is complete.
This project was an early development of what we now call LLM agents. It is a collab with Eddy Liang during the summer of 2023.
https://www.youtube.com/watch?v=3ZKDWljOhm8
</aside>
<aside>
📍 MindMap Extension (unfinished)
frontend
I wanted to create a ChatGPT tool that visualized your chat in real time. This visual tree allowed users to see the context and quickly navigate (using Vim-like controls) to edit, copy, or expand on specific chats.
I got up to the point where nodes were automatically created (still a bit buggy), but haven’t finished the project.
I also couldn't find anyone else who wanted this tool other than myself — but that’s enough reason to pick up the project again and finish it.
some original mockups
</aside>