Essays and notes on AI, GPUs, and software I am building.
Vireeel shipped to the App Store today. The honest version: this is a crowded corner of the store, I spent four months convinced I could out-build the incumbents, and I couldn't. So I made almost all of it free, fixed the bugs that mattered, and shipped. Here is what it does and how it actually got out the door.
The biggest Kekasatori release yet. It still turns a link or a paper into something you can study, but 1.4 adds the machinery to actually build: a Cockpit that makes a Workspace the unit of work, a compute dial that runs the same job locally or on a rented GPU, reactive marimo notebooks with SQL, a learn-by-implementing Reinforcement Learning track, and an agent that remembers you and has to ask before it touches anything. Still no account, still your keys, still local-first.
The engineering story behind Kekasatori 1.4. How three loosely-related pillars became one Cockpit built on a four-object model, how one ComputeBackend protocol lets the same job run locally or on a rented GPU, how the agentic memory works (a hashing embedder, multi-signal recall, and a temporal knowledge graph that closes facts instead of deleting them), why the agent has to ask before it acts, and the AppKit tracking-area bug that made a text cursor haunt every tab.
The biggest Kekasatori update yet. It still turns a link or a paper into something you can study, but it grew three new pillars for the people who actually build and evaluate models: Discover a literature across five scholarly sources, Compare any models side by side, and Run real code and benchmarks inside containers. Plus a deep-research agent that reads full papers and argues with itself, a learn-by-implementing track, and an MCP server so your own AI assistant can search your library. Still no account, still your keys, still nothing leaving your Mac unless you ask.
Since 1.2 the centre of gravity moved. Kekasatori is becoming a cockpit for people who build and evaluate models, not just study them: a multi-source literature Discover with an in-app paper page, a side-by-side model Compare arena spanning local, cloud, and Hugging Face, and a Run tab that executes code and benchmarks inside containers. Plus the design decisions, and the two features I built and then deleted.
The latest Kekasatori update adds a decentralized community wall (Nostr for the index, IPFS for the files), a paper-to-podcast generator, a Shu-Ha-Ri mastery board and focus timer for actually working through your study packs, and a calmer redesigned home, all without giving up the no-account, local-first principle.
A long look at how Kekasatori is built, from its SwiftUI and SwiftData architecture to on-device retrieval, chunking, and the decision to drop raw YouTube streams for the embedded player.
A local-first macOS study app that turns videos, papers, and articles into transcripts, flashcards, mind maps, and a grounded AI tutor.
A practical walkthrough of how Flash Attention reduces memory traffic and speeds up transformer training.
Notes from learning CUDA memory hierarchy, occupancy, and writing my first custom kernels.
