Portfolio

Open CodeTree

A LLM chat powered fork of the CodeTree project which allows running Typescript and Javascript in browser with a sandboxed environment with ESBuild and WebAssembly. Rapidly prototype in the browser!

React
LLM
WebAssembly

Demo Code

Telephony AI Chat

A telephone based voice assistant prototype built on top of Websockets and Twilio using state of the art voice and text models. Demo available upon request.

WebSockets
Node.js
OpenAI
Deepgram
ElevenLabs

Code

Chat with Internet

A Large Language Model functions powered chat with browser based voice to text transcription. Functions call APIs like Google Search and Sports APIs.

OpenAI Functions
Next.js
Transformers.js

Demo Code

Apollo OpenAPI

Generates an OpenAPI spec from Apollo graphql projects with the standard Swagger UI docs page. Bringing the worlds of REST and Graphql together to make APIs more accessible for LLMs.

Apollo
Typescript
Express

Code

Edge chat

This prototype explores using browser based embeddings to generate an in-memory vector store combined with a LLM for a more private and cost effective Chat with PDF experience. More to come with this soon!

Apollo
Typescript
Express

Demo Code

CPU based LLMs

Running compressed large language models on CPU is cost effective and rapidly improving. Most recently I have quantized the Magicoder 7B which scores near the top on the CanAiCode leaderboard.

Docker
llama.cpp

Demo Code

Open Source Image LLMs

Image based language models are a new frontier in terms of creating unique AI experiences. Currently most users are only exposed to paid closed source models, but models like LLaVA are very competitive with the benefits of being run your own infrastructure.

Docker
llama.cpp

Demo Code

Web LLM Embed

This prototype explores combining an WebGPU powered LLM with vector store embeddings for a document based chat. With smaller language models becoming more powerful this may be the archetype for future applications.

WebGPU
Transformers.js
React.js

Demo Code

Backseat Pilot

Coding co-pilot extensions are a dime a dozen now. This was one of the first to support your own local open source LLM where the model fills the code rather than auto suggesting completions. Plus it is designed to be a project you can fork and customize to your own needs.

VSCode
ESBuild

Demo Code

Google Clone with AI

Building a Google clone is a fun frontend exercise, but adding OpenAI adds an interesting dynamic. As end-users we want to choose our Search Engine and AI provider. So instead of using Bing with ChatGPT, or Google with Gemini you can create your own.

Next.js
OpenAI
Google Search

Demo Code

Dishcovery.eats

An OpenAI functions prototype exploring using the SerpAPI restaurants API. Uses Yelp and Google with simple search UI.

Next.js
OpenAI Functions

Code

Buildless Blog

Typical modern javascript projects require an installation and build process. This project uses alternative libraries to remove this complexity including import maps. Not practical but a fun experiment.

React.js
Import Maps

Demo Code

Signifier

This was a fun attempt to create a command line tool for calculating significance in an A/B test with error bars. I converted a statistics processing script I had learned with Python and converted it to work with Node.js.

Node.js
Plotly

Code

Apolloskine

Creating a to-do list app is a rite of passage for any developer. I wanted to put my own twist on it to learn more about Apollo with a touch of skeumorphism to mimick the a notebook. Included a novel way to share short notes via links.

Apollo
React.js

Demo Code

Skymatt

I wanted to create a simple webpage that mimicks the color of the sky based on the time of the day using only CSS. I ended up having to write a little javascript but definitely improved my CSS chops in the process. Kind of a meditation app!

Demo Code