Still working on Hopp, an oss pair programming app. A few months ago we ported part of the frontend in rust, to work around the webrtc limitations imposed by webkit (Hopp is written in tauri), which caused some stability issues. Now the app is more stable again but we are still making improvements and fixes.

https://github.com/gethopp/hopp

I'm continuing to expand my own deep learning library [1] built with numpy-primitives to support LLM post-training techniques like supervised fine-tuning (SFT) and reinforcement learning with GRPO. It's a good learning experience to work without all the high-level abstractions to "build a wheel" and "use that wheel to build a car".

I'm also looking into coding harness self-improvement [2]. An inner LLM (raw LLM request) + harness solves coding tasks, an outer agent like Claude or Codex that proposes harness changes. I experimented with many things in the past few months that made me realize this self-improvement thing that everyone is talking about is just an experiment design problem. I wrote about it here [3]. I'm continuing to improve the infra around the self-improvement loop, to increase signal-to-noise ratio per experiment. I'm also generalizing the infra to expand beyond terminal bench tasks and to collect some data across different models (harness-bound vs model-bound).

[1] https://github.com/workofart/ml-by-hand

[2] https://github.com/workofart/harness-experiment

[3] https://www.henrypan.com/blog/2026-05-25-self-improvement-ha...

I've continued working on Eidetica, my decentralized database project. I recently added support for a client/server architecture so that it can be transparently run as a local daemon for background sync and sharing the local storage with multiple users. I've been making progress on integrating blob storage next, as well as scoping out WASM based "lenses" for handling decentralized version/schema updates. https://github.com/arcuru/eidetica

I've primarily been testing it by building out my AI tool chaz into an Eidetica-native AI Agent framework for decentralized Agent sessions. It's working surprisingly well, it maps pretty well onto the storage model and it's uncovering issues with Eidetica I need to fix (which was always my primary reason for building it anyways). https://github.com/arcuru/chaz

Separately I'm building OptiMap, a SIMD-accelerated hashmap repo that explores the design space for hashmaps and benchmarks different approaches. This is mostly for my own learning but I'll eventually turn into a blog post. https://github.com/arcuru/optimap

Working hard to finally ship my co-founded product - Edit Square, an online motion graphics editor https://editsquare.com/ - private invites are going out this week, wish me luck!

This has been in the works for many years! The project originally started as web forms driving After Effects templates on a Windows server, and has now evolved to a point where the web technology landscape has matured enough to build a full-on motion graphics editor right in the browser, using WebGPU and WebCodecs.

I'm working on an iPhone app to run iOS-native agents using both cloud and local models, powered by llama.cpp. It has access basic iOS tools such as calendar, reminders etc. but also more advanced ones like a custom JS environment running on QuickJS that can use various custom modules like an HTTP client, Git etc.

It's a project I have been working on for quite a long time and I released it on TestFlight about a week ago. It was really nice to work on something end-to-end, from creating a wrapper around llama.cpp with support for prompt caching/forking and automatic model loading and unloading based on device memory constraints, to the custom agentic harness the app runs on. I have also spent quite a lot of time on agent execution modes that I hope can enable to more easily reason about agent security regarding prompt injection attacks.

What I'm really hoping for now is to get actual feedback, to know if users end up having real use cases where the app is truly useful / interesting for them, to understand what should most urgently be improved etc.

https://bilembi.app/

Yesterday I tried to take another crack at reverse engineering my e-scooter. I am don't have the energy to sit down and read through the decompiled android app, so I thought I'd try using Qwen3 Coder next for that but it kept looping when analyzing it, probably due to my terrible prompt. I tried with Mistral Vibe and then OpenCode. OpenCode showed me that the agent was trying to search for all kinds of strings over and over again. I am thinking I'll try yet again this week

I am close to buying a Claude sub but the thought of it going haywire and costing me extra money in tokens is too scary yet. Not to mention how much provider LLMs (not sure on the correct terminology for them as opposed to local) could hamper your reverse engineering efforts (looking at you Fable).

I really want to solve this scooter and have my own app for it. There's a firmware update feature in the app, maybe I could dump the firmware at least and that would help. Anyone have suggestions on what language model would be the best for this task (analyze decompiled and obfuscated android app / analyze dumped firmware ) ? i have a 96gb macbook and would prefer a local one (I guess to make myself feel better for having spent money on it?) but something through OpenRouter or whatever would do just fine as well

Im putting together a foraging map which correlates wood, heat, temp, and whatnot to tell me where various edible fungi fruit. It is mostly working so far, has bike trails and we tested it at least once.

The most challenging part was getting MVTs to fly but it is very fast already even in mobile. The fun part is tarring the solver solves correctly :) no public version though but I can upload a screen grab somewhere should anyone be interested.

I have been going regularly to the Edinburgh Festival Fringe and their app is always terrible. Connectivity is very spotty since the city is swarmed with way more people than usual and the older buildings can kill your signal.

So I've been working on https://fringeflypost.com/, an event tracker with maps, search and filter, scheduling, and sharing with friends that's offline first. It syncs down a locally stored sqlite database and caches assets pretty aggressively.

(You don't actually need to sign up, and you can just jump into the list of shows directly here https://fringeflypost.com/shows).

I am working on two things:

- SophAI (https://www.sophai.app/): an app to connect the dots across cross-domain. As a CTO, I read across multiple domains (tech, design, business, e-com) and often have to connect the dots. I am building this primarily for myself. It is basically a rss parser with a big AI prompt to connect the dots across the blog posts. As I type this, I'm working on adding podcasts to the app.

- CTO field notes (https://www.ctofieldnotes.com/): collection of essays growing out of my 30 years in IT services. One essay every Tuesday.

Working on Authbound. Digital identities are slowly becoming more prominent. With EU mandatory regulations, it will just boost this market. So we're building eIDAS 2.0 / EUDI Wallet Plug & Play SDK integration for businesses that don't want to figure out the compliance and technical mess themselves.

Side effect: KYC gets stupid cheap. Cryptographic credential verification vs. traditional document checks is not even close on cost (≈90% cheaper). Qualified Electronic Signatures and EUDI Wallet Payment systems are also coming in the following years.

https://www.authbound.io/