Builder's manifest, rev. 2026
Gianni Sanrochman
I just love building. Lately: a freight company, an inference engine, an operating system. I own the problem, the architecture, the verification, and the pager.
I founded and run Aircrate, a freight forwarder connecting the USA and China to Suriname, on a stack I built end to end. I maintain pulsar, an open-source Rust and CUDA engine that streams 743B-parameter models from SSD through two consumer GPUs. MSc in Psychology from Erasmus University Rotterdam, interdisciplinary track on digitalisation in work and society.
Cargo on board
Aircrate
Freight forwarding from the USA and China to Suriname. Storefront, quoting, multi-rail payments, tracking, operator tooling. Runs on three-tier high-availability PostgreSQL; the last line of defense is a Raspberry Pi.
pulsar
SSD-streaming inference engine for giant mixture-of-experts models, in Rust and CUDA. Runs 743B-parameter models on two consumer 16 GB GPUs. Benchmark-gated development, 200+ stars.
neutronstar
CUDA evolution of antirez's ds4: io_uring expert streaming, cross-layer prefetch, and the first MTP speculative decoding for GLM 5.2.
protostar
Continual-learning research: a 134M-parameter model that retains 77% of learned skills across sequential consolidations, via replay-gated memory.
encoder-free-speech
Speech research on Gemma: audio understanding and generation without a separate encoder stack. Paper writeup in the repo, WER-gated results.
agora
A Rust MCP hub that links AI coding agents across machines and harnesses, with exactly-once message delivery. My agents use it to work as a team.
Method
Everything above ships the same way: iterative and measured. Held-out benchmarks decide what lands, controls are built in by default, and negative results stay in the record.