Aayushya Patel · Cupertino, CA
_{Founder of Kahlus · Founders Inc. Canopy S26 · Stanford SIMR '26}

site · kahlus-bench · kahlus-skills · recall

 ·   ·   · 

Neural forecasting, and the question of when a result is real. Most of what I build is a benchmark, a bound, or a negative result I could not talk myself out of.

kahlus-bench · neuroforecast — Leakage-sealed benchmark for neural-coupling claims, and certified directed-information graphs for multi-organ coupling. Finite-sample certified, not asymptotic hand-waving.

Recall — Local second brain. One search box over files, Gmail, Calendar, Drive, Canvas, Notion. Sub-200 ms semantic search, nothing leaves the machine.

kahlus-skills — Agent skills for research artifacts a skeptical PI would believe. One rule across all five: every number reproduces from a named command, and scope is stated honestly.

hil-failure-modes — Six concrete failure modes where AI-generated embedded firmware compiles clean and fails on real silicon.

Also: mentra-eeg-miniapp (live EEG-derived cognitive-state readout on smart glasses) and bambu-studio-safe-updater.

Two repos below are private. I'd rather name them and say so than link a 404.

Kahlus-V1 — Leakage-controlled EEG forecasting harness, the NeuroTwin research engine. Six evidence gates, and M3 and M4 are committed as gate_passed: false. The failures ship with the repo, because the point of a gate is knowing when the result isn't real.

eeg-forecastability-limits — How far ahead can you forecast raw scalp EEG? On BNCI2014_001 the standard task builds a target that shares 127 of its 128 samples with the input, so the score is mostly copying. Rebuilt as a single strictly future sample: raw voltage reaches r = 0.885 at 4 ms, 0.176 by 256 ms, and is indistinguishable from guessing by ~1 s. Band-power envelopes are not — gamma still scores r = 0.44 at 8 seconds. It also carries the correction to my own earlier README, which turned out to be measuring my regularisation choice rather than the signal.

Happy to walk it through on request — email.

• Distrust green CI. The tests passing is a claim, not a result.
• If a baseline got 60 steps and mine got 400, I didn't win.
• Publish the failed gate. An honest smaller artifact beats an impressive-looking fake one.
• Agents as adversaries: propose framing, ask for the strictest possible critique, adopt, repeat.

_{IEEE ISEC 2026 (sole author) · ASME IMECE '25 · SCCUR '25 · NCUR '25, '26 · BMES Bay Area '26
email · aayushya.net}