A small, self-hosted programming language. The compiler is written in Whispem itself, compiles itself, and produces byte-identical output with the reference Rust implementation — a stable fixed point, verified by a reproducible bootstrap. It runs on a standalone C VM with no dependencies beyond a C compiler.
Self-hosted compiler — the full pipeline (lexer, parser, codegen) written in ~1700 lines of Whispem. Source in, bytecode out, identical to the Rust output.
Verified bootstrap — the compiler compiles itself, and both outputs share the same hash.
Standalone C VM — a single-file runtime (~2000 lines), 34 opcodes, interactive REPL, and a disassembler.
204 tests, zero warnings — Rust tests plus autonomous C VM tests with bootstrap verification.
A distributed key-value and object store in Rust, built entirely in the open — starting single-node, then rewritten as a full distributed system. Focused on understanding the system rather than chasing raw numbers.
Raft consensus for cluster coordination.
Two-phase commit for distributed transactions.
Write-ahead log for durability.
Virtual sharding for elasticity, and an S3-compatible API.
Rust · distributed systems · consensus · storage
sussurro.cpp — offline neural translation, from scratch
An offline voice-to-voice interpreter across English, Spanish, French and Italian. I reimplemented the inference of Marian / OPUS-MT neural translation models from scratch on ggml — encoder–decoder Transformers running entirely on-device, with no server and no network at runtime.
Marian / OPUS-MT inference reimplemented on ggml — encoder–decoder Transformers, greedy and beam-search decoding, an incremental KV cache, quantized weights.
Twelve translation directions, no pivoting — one multilingual model for Romance ↔ Romance, bilingual models for English ↔ Romance.
Voice in, voice out — speech-to-text via whisper.cpp, text-to-speech via sherpa-onnx.
Native desktop app — a Tauri UI wrapping the whole pipeline; every stage also works from the CLI.
A networked key-value store written in pure x86-64 assembly — no libc, syscalls only. Built to see what a server looks like with every abstraction stripped away.
epoll event loop — 200+ concurrent clients on a single thread, non-blocking, with partial lines carried across reads.
Hand-written hash table — FNV-1a, open addressing, tombstones that get reused so delete/insert cycles never rot the table.
mmap memory arena — mapped once, handed out byte by byte.
SIGPIPE-proof — a client that dies mid-write can't take the server down. Compiles to a 13 KB static binary.
A terminal-native data profiling tool in Rust — htop meets pandas-profiling. Explore multi-gigabyte datasets instantly without leaving the shell. Built with Polars and Ratatui.
Rust · data tooling · Polars · TUI
More projects — including a few written purely for the joy of it — live on GitHub.