projects

cpu inference for ternary neural nets. weights are -1, 0 or 1, so the matmul is adds and subtracts and no multiplies at all. hand written simd, no third party dependencies.

point it at a repo you have never seen and it tells you the stack, the shape of the tree, and what looks wrong.

shared memory ipc with no copies. one writer, many readers, a ring buffer in mapped memory, nothing serialised on the hot path.

rotation based memory that does not decay. a recurrent cell that swaps gating for rotation on a complex valued phase manifold.

memory. or a brain. still deciding which.

voice agent that remembers you between calls.

a c++23 callable wrapper covering the same ground as std::function, with the storage, dispatch and lifetime written from scratch. inline buffer so small targets never touch the heap, per type vtable, move only, every signature qualifier preserved.

layer 4 firewall in ebpf and xdp. filtering happens in the kernel at the xdp hook, before the packet reaches the network stack. dynamic blacklist, rate limits, tui on top.

membership that refuses to lie. it hands back a probability instead of a boolean, because a gossip layer does not actually know.

an ide that runs entirely in the tab. the toolchain is wasm, web workers are the processes, indexeddb is the disk. no server anywhere.

context engine on top of helix. 2m parameters, generates text on cpu, no gpu anywhere.

chrome extension that tells you the font under any text you select.

Inference-time reasoning framework. Trade compute for correctness.

a proxy that speaks the postgres wire protocol directly. logs queries, routes reads away from writes, rewrites what it is allowed to and blocks what it is not.

finds the code that disagrees with the rest of your Go codebase. Infers lock and nil-check rules from the code itself, engler-style.

finds sybil farms by fingerprinting the order of every transaction a wallet has ever made, then clustering the wallets whose orderings match.

writes the commit message from the diff, so the log stops saying "fix".

my blog. long posts, papers i implement, books.

GPT built from scratch — assembled from my NeetCode ML course submissions

dsa slop

make your agents faster a lil bit

an inference engine for a 200b mixture of experts model on a 16gb laptop, where nvme sits on the critical path of every token and not just the cold start. two mechanisms: predicting which experts are needed several layers ahead, and scheduling prefill around experts rather than around layers.

dither ui components. react, no dependencies.

four stl containers written from scratch in c++23. optional, vector, variant, unordered_map. real allocator support, manual lifetimes, exception safety.

points at a javascript repo and finds what nobody deleted. dead files, unused exports, components nothing renders, imports that go nowhere.