Projects.

MemePad

A memecoin launchpad on Solana, with an ape called Razz. 2024.

A launchpad for memecoins on Solana. Max ran the business side and I built the technical side with a small team: the Next.js front end, the Go backend and the Telegram bots around them. The presale opened in April 2024 and the public raise ran from 15 to 18 June.

It reached 42,000 users and raised $1.1M in under three months. A few days after the raise closed, it was hacked. One of the developers rugged it. The lessons were expensive. I wrote about 1,800 of the 3,600 commits across the org.

The brand was a world more than a logo. Razz, a grey-haired ape who bought his way out of a Texas bitcoin mine, ran the MEMELab where launches were cooked up. Every project on the pad was meant to arrive as a new character in that story. One animated short got made.

Token holders got the MEMEVerse Members Pack: a free NFT mint in five tiers, Bronze to Diamond, each card animated and stored on Arweave.

Neotrade

Social trading on Solana: clans, competitions and rewards, inside Telegram. 2024.

A trading platform where trading was a team sport. Traders joined clans, clans traded together and paid part of their winnings into a shared treasury, and whole clans went up against each other in trading competitions called Escape the Matrix. Every trade earned points toward airdrops. It started as a Telegram bot and grew a web app on top: charts, a trending list, a portfolio.

The whole backend was written from scratch in Go, including the trading engine and the swap execution under it. Go had almost no Solana tooling then, so the on-chain indexer was ours too: decoding every instruction, parsing swaps, following inner instructions and inner transfers through each transaction.

The brand was The Matrix. A white rabbit for a logo, and Razz from MemePad in the chair.

The first version shipped with quests for airdrop points and brought in 20,000 users in three weeks. I wrote 494 of the 721 commits on the app, July 2024 to August 2025.

ChatGod

A multi-model AI assistant on Telegram and WhatsApp. 2023.

ChatGPT, image generation and voice inside Telegram and WhatsApp, including group chats. Send it a voice message and it answered in voice. Forward a friend’s voice note and it transcribed and summarised it. Send a YouTube link and it summarised the video. Ten free messages, then a subscription.

The first commit is from 7 February 2023, three weeks before OpenAI opened the ChatGPT API and four months before it shipped function calling. The model took plain text in and gave plain text out. It could not pick a tool. So we wrote our own protocol for it to ask for one, in a format we could parse reliably, and an event-driven backend with a router that sent each request to the right model: OpenAI, ElevenLabs for voice, Stability for images.

Contractors had started on it a few days before I was hired. I threw their code away and rewrote it from scratch in Go, then hired and led a team of three as CTO. 377 of the 467 commits are mine.

At its peak it had 200,000 users and about $1,500 a day in OpenAI bills. The plan beyond the bot was a marketplace for AI agents, before agent launchpads existed. It died of management decisions and product-market fit. We were early, and we should have stayed with it.

Ishizue

A Rust framework for microservices. In private development.

I have written backend services for a long time, and most of them repeat the same work: discovery, transport, retries, config, logging. Pulling that out into something shared is how I ended up in microservices, and then in the frameworks for them.

Go was my main language, for the performance. I became one of the maintainers of Go Micro, a framework with 23,000 stars, and I am the second-largest contributor to its plugins. Then a few of us started designing the next version: a better developer experience, cleaner code, faster at runtime. We called it go-orb and built it in Go, by hand, before ChatGPT. It was never finished.

Ishizue (礎, foundation stone) takes the same principles into Rust, built with AI. Define the API in a proto file, generate the code, and write a handler that doesn’t know whether the request came in over HTTP or gRPC. Discovery, retries, circuit breakers and observability are plugins, and a handler wired up wrong fails to compile instead of failing in production.

About 400,000 lines across the core and the plugins so far. Tensai runs on it.

SuperFans

A SocialFi launchpad on Solana. October 2023 to April 2024.

Communities did tasks for points, points turned into airdrops of the $FAN token, and the launchpad listed tokens from creators.

A small team built it: a Next.js front end, a Go backend on Postgres, an Anchor program on Solana and an indexer. I wrote 254 commits, 163 on the front end and 90 on the backend.

There was a reputation score, staking, airdrop seasons, and a token generation event on Raydium on 10 January 2024. The marketplace code outlived the product: re-skinned, it became the MemePad airdrop marketplace.

The project ended in an ugly co-founder divorce. I was merely one of the children.

Agents and their harnesses

How I work now, and the tooling I build for it.

Most of my time now goes into AI agents: building them, giving them tools and access, and finding new work for them to do. Thousands of hours so far, across different systems.

The models are smart. That was never the hard part. Anything worth doing with them is complicated: a large codebase, a research project with years of data behind it, a process with real money running through it. Hand that to an agent the usual way and you get a mess that looks like progress.

The difference between that mess and excellent work sits almost entirely in the setup around the model. What it can touch, what it knows, how it has been taught to think. Mine get philosophy lessons. There are a hundred small decisions in a setup like that, and most of them only show themselves once you have got them wrong.

The part I put the most care into is specialists. Each one carries the working method of the people who are best at a craft, taken from how they actually work. They sit in on most of what I do in a day, and between them they carry work that would otherwise need a whole team.

Doing this every day shows exactly where the existing tools stop, so I build past them: an agent workspace, the search my agents run on, and the framework under both. Almost anything is possible with these systems, once someone who knows where they break has set them up.

Hardcut

A control room that runs a live show from one screen. In testing.

The producer picks the show, preps the running order, and during the hour fires what goes on screen: a headline, a post, a slide, the cast, a segue into an ad break. OBS takes the whole programme as one transparent browser source.

Each show has its own look, and four layouts move the cameras around whatever is on screen. Stories come in from a newsroom that pulls posts from X. Prices come from a live DexScreener feed, read through a decoder for its binary format.

First commit on 9 September 2026. 217 commits in eight days, about 26,000 lines of TypeScript.