This is a list of notable projects that I’ve finished and or currently maintaining. My (considerably longer) list of unfinished ideas can be found here.
Portal – zero-config P2P encrypted folder syncing
Portal is a command line tool that syncs folders between multiple devices. Perfect for syncing photos/videos/code between many devices without using a 3P tool like GitHub, Email, or Google Photos. Built on top of the Hypercore protocol with Ink for the CLI.
Built with: Hypercore, React, Ink, and Typescript
Quartz – create and publish digital gardens for free
Quartz is a tool and workflow to make maintaining and publishing a digital garden and second brain extremely easy. Think Obsidian Publish but free. Links are scraped and processed using a GitHub action written in Go and dumped into a YAML file to be rendered by some prefined templates with Hugo.
Legist – a platform to summarize policy for democracy
Legist is a web platform that allows users to digest policies in an efficient and accessible manner. Legist allows users view automagically summarize pieces of policy + legislation while still maintaining the key takeaways, view and filter policies by category, and subscribe to periodic rollups on updates. Frontend was built with React + Typescript + Chakra UI. Text summarization was done using DistilBART, Named-entity recognition with BERT, and zero-shot text categorization using BART. All models were served with BentoML. Built at Hack the North 2020++, winning the Founder Institute Fellowship Prize and finalist among over 3000+ participants
ctrl-v – a modern, open-source pastebin
ctrl-v is a modern, open-source pastebin with LaTeX and Markdown rendering support. Any user can create a paste without an account, with the ability to protect it with a password and set an expiry date. Additionally, ctrl-v does code highlighting as well as LaTeX and Markdown rendering. Pastes are stored in a MongoDB Atlas instance. Backend is a containerized Go service deployed on Google Cloud Run. Frontend is a SSR Next.js app hosted on Vercel.
reflect – a mindful website blocker for the productive
reflect is a browser extension with 800+ active users focused around asking users to reflect before visiting distracting sites, helping to reduce mindless scrolling while still being able to get work done. During closed-beta, we created a Go service that logged user intents to a Cloud SQL database and did intent classification by serving a basic Flask API. We then trained an LSTM network in Keras on the closed-beta data and augmented it using NLP data augmentation techniques, reaching ~86% classification accuracy. Finally, the model was ported to Tensorflow.js where it runs in-browser within the extension which is written in Typescript.
Built with: Kubernetes, Docker, GKE, Cloud SQL, Keras, Tensorflow.js, Flask, CircleCI, TypeScript, Python, Go
nanoDB – a simple, easy, and debuggable document database
nanoDB arose out of many frustrations I’ve personally come across while prototyping, namely 1) difficulty of debugging data 2) faffing around with language specific drivers and 3) reference resolution. As a result, nanoDB stores everything on disk as a JSON document, has built-in reference resolution, and can be used fully through a REST API. Think of it like Redis but with MongoDB style documents — all of which is on-disk, human-readable, and through a REST API. The project is fully written in Go and is thoroughly unit-tested. It features a standalone server binary which creates a nanoDB server, as well as a shell which allows you to do some basic document inspection.
Built with: Docker, Go
readAR – an AR app to help those with learning disabilities
readAR is a mobile AR app re-renders text to be more dyslexic-friendly, and adds context-dependent word definitions and images. A custom BERT model was created and trained for word sense disambiguation (WSD), achieving a 76.6% F1% score on the test dataset which is only ~5% away from state-of-the-art. This model is served through Flask on an Oracle VM Instance. The API is also responsible for our image processing pipeline, which is a conglomerate of different Azure APIs (OCR, Text Analytics, Bing Search).
Built with: PyTorch, Flask, Docker, Azure, Oracle Cloud, Python, Bash
Impostor – group pomodoro timer with a twist
Imposter is a productivity timer designed to keep friends on task together even when working remotely. The Chrome extension monitors your browser tabs, checking against a blocklist of unproductive sites. If one of those websites is visited, the backend will be notified via Firebase and will notify all users in the room through a WebSocket connection. Chat and character customization is also supported.
Speech2Braille – a wearable device to transcribe speech
Speech2Braille was created to help the over 360 million people in the world who have debilitating hearing loss. This project entailed creating an end-to-end speech recognition system using an Deep LSTM and a portable device to display braille. The device is able to recognize audio and transcribe it into Braille through the haptic feedback device via a novel neural network architecture. The feedback device is a self-made GPIO hat, consisting of 6 solenoids. The neural network itself is 2 layered LSTM-CTC network with 256 hidden cells in each layer, achieving 92% state-of-the-art word error rate on the TIMIT dataset.
Built with: Tensorflow, numpy, Raspberry Pi, Python, Bash
PacketBook – blockchain banking without internet
PacketBook is a financial accessibility chatbot, fully accessible through the SMS (text messaging) protocol. Users are able to issue simple commands to register, check their balance, deposit, withdraw, and send money. These commands are secured with two-factor authentication through a Flask server on Heroku with Twilio. The backend that handles transaction is written with Node and Express and deployed on stdlib. PacketBook is unique in that it leverages the Stellar blockchain and tokens (XLM) for its transactions which greatly reduces operating overhead with its minimal transaction costs (around 1/100 of a cent per transaction).