It's Time to Build

A proof of concept for building a type-safe Next.js application with tRPC over Hyperswarm, Holepunch’s peer-to-peer networking layer. The demo is a health-question chat with an AI assistant called Murphy.

Requests and responses travel through a pair of peers and local Unix sockets, keeping the transport separate from the application logic. The goal was to use familiar web tools for a peer-to-peer app without making the rest of the code depend on Bare’s runtime limitations.

An AI food tracker built as a Telegram Mini App. Describe a meal in a message, send a voice note, or take a photo, and Moyasu estimates its calories and macronutrients.

You can correct ingredients, quantities, or estimates through conversation before adding the meal to your food log. The log is accessible from Telegram or a browser, making tracking less dependent on filling out forms and searching for individual ingredients.

A Slack assistant that makes a team’s existing knowledge easier to find. It uses language models and a vector database to search the meaning of messages and attachments, rather than relying only on matching keywords.

The idea was to help new teammates get up to speed and let people ask questions where the conversations already happen, instead of repeatedly asking colleagues to locate old discussions or documents.

A Notion task database designed to handle recurring work while keeping a record of what has already been completed. It preserves task history alongside the next occurrence, so repeated tasks remain useful both as a schedule and as a log of past work.

A Python library that adds Django REST Framework-style input validation to Graphene GraphQL mutations. It supports checks on individual fields as well as checks that depend on several fields, and follows nested inputs all the way down to their scalar values.

Validation errors include the path to the offending field, so a client can show the error in the right part of a form. Validators can also transform values before they reach the mutation, such as replacing an identifier with its corresponding object.

A web project collecting anonymous stories about bad recruiting experiences in the technology industry. It gave people a place to share the frustrating, strange, and sometimes absurd interactions that usually stay buried in private conversations.

A Chrome extension that replaces invisible zero-width characters with visible emoji. Those characters can be embedded in otherwise ordinary-looking text, so the extension makes them easier to notice when reading or copying a page.

I built it as a small experiment after reading about using invisible characters to hide information in copied text.

A collection of my early electronic music experiments and releases. These are the tracks I made while exploring electronic production, collected on SoundCloud.

A network discovery tool that listens for broadcast ARP packets to identify devices and their IP and MAC addresses as they announce themselves. It uses passive packet capture rather than repeatedly scanning a range of addresses.

Built for Istanbul Packathon 2016, where it won first prize. It supports callbacks and middleware for reacting to discoveries, such as logging a device or sending a notification. Like other local discovery tools, it cannot see peers hidden by Wi-Fi client isolation.

A small Android proof of concept that uses a phone’s barometer to approximate pressure-sensitive touch. On a waterproof phone, pressing the screen can change the air pressure inside the enclosure; the demo uses that reading to simulate an interaction like iOS Force Touch without a dedicated pressure-sensitive display.

Built with Cordova and a barometer plugin as a quick experiment. The project was covered by Gizmodo.

A small Linux utility for sampling hardware performance counters at very short intervals. Built on PAPI, it can attach to a process, monitor a chosen set of events, and export measurements as CSV for analysis.

I built it for experiments that needed finer timing than perf stat offered. In some tests it reached a sampling resolution of three microseconds, making it useful for observing brief changes in instructions executed, cache accesses, and cache misses.

An article about treating production feedback as a deliberate part of software development. It makes the case for testing and observability that help teams understand how a product behaves in real use, rather than treating production as an afterthought.

My PyCon Balkan 2018 talk about importing more than 10 million images under demanding time constraints. It covers the work at Hipo Labs that combined AWS Lambda and DynamoDB to make the import process 10× faster.

That improvement helped the client work with more dataset providers. The talk follows the engineering decisions behind the speedup and the practical constraints of processing images at that scale.