bio

i'm 21. built my first pc at 11. homebrewed my 3ds for pokemon nuzlockes.
hex-edited games to get outside the map.

rules are silly. nothing's impossible till physics proves it.

technion pre-engineering, left after a missile hit campus. economics undergrad at binghamton.

taught myself ml from scratch, december 2025.

spent six months on hardware-enforced alignment.
pivoted, it only served closed labs (safety is not control).

now: building an open source ai efficiency lab.
frontier ai, radically cheaper to train and run.
science shouldn't be this expensive.

interested in joining me? reach out.

conceptual builder, not traditional coder. i design, my agents work.

founding engineer at duino, now technical advisor.

upstream

projects

• hypertok the fastest tokenization on the web. exact token ids in the browser and node, 6.2-42.9x the fastest incumbent per workload, up to 295x js-tiktoken. benchmark harness ships in the repo. race it live, or on npm.
• turboadam drop-in adam replacement. 9.845x smaller optimizer state than adamw, 6.50 bits per parameter. +1.224% on tinystories, -0.094% on wikitext-103, matched 500-step gpt-2 124m runs. on pypi.
• dcaf and noosphere interpretability tools. circuit discovery using fine-tuning as a controlled perturbation, and a 3d explorer for embedding spaces.
• arrow field embeddings keeps what a 3d projection throws away, encoded as arrow direction and magnitude. benchmarks in progress.
• easywheels prebuilt cuda wheels, no 30 to 120 minute source builds. retired when uv's team started building the same thing.

research

• for the lab, in development an ffn that skips about 85% of per-token ffn compute, within about 6% of standard in small-scale tests. a low-bit residual master weight at about 89% perplexity parity, 6.75 bits. optimizer infrastructure that generates its own kernels. a ternary training format.
• navier-stokes an enstrophy-weighted directional fisher information regularity criterion. conditional, not a proof of global regularity. many rounds of ai red teaming, no human expert has checked it.
• entropy-attention and r-attention two attention mechanisms i designed and tested. neither beat standard attention across an 8-task battery. published anyway.