Research
-
Quantization in diffusion vs autoregressive LMs
Do masked diffusion LMs tolerate low-bit weights better than autoregressive models at matched scale? Matched pairs, thresholds set before each run, all on one 8 GB GPU. At INT4 the diffusion model degraded about half as much.
-
Recurrent depth on a pretrained 4B model
Seven attempts to retrofit a looped core onto FrogNano-4B, none better than a compute-matched LoRA. A training-free gain on HumanEval turned out to be reproduced by a noise control.
Open source
-
von
A small encoder model that picks between options or scores them in one pass, without generating text. Runs on CPU, with Python and TypeScript SDKs.
-
rubick
Chat, embeddings and decisions served from one local model instance.
-
vole
A Linux system cleaner and monitor, ported from Mole.
Products
-
hellow
Unified chat for streamers across Kick, TikTok, Twitch and YouTube. Free plan and a Pro subscription.
Work
Backend and platform engineering in finance, aviation, public sector and ticketing: a ticketing platform with 1M+ monthly users (staff engineer, Ingresse), a regulated FX system (tech lead, Citi), and aviation communication systems. M.Sc. in Applied Computing (UFPA).
I also take on consulting work.