川井 涼介
Entrepreneur, Engineer
I'm a 15-year-old high school student from Japan, now at a high school in Toronto.
Projects
67% of the facts for 6.3× less GPU energy. A reader shown only a 32k-token correlation diagram of a 4-hour, 179k-token agent session answered 63 of 96 questions. The same reader given the whole transcript answered 94 using 6.3× the energy (Qwen3.8-27B, Sep 2026).
Python · LLM · Attention
Claude Code that doesn't forget. Sanma gives Claude Code a memory that persists across sessions, compactions and restarts.
Python · Developer Tools · Memory
Hackathons
Youth Code x AI · June 2026
You control a 3D dashboard with your hand through a laptop webcam by pointing, hovering and pinching. It needs no mouse and no special hardware.
AI Tinkerers Global Hackathon · September 2026
A player types a request in Minecraft chat, and an agent builds it in the live world. A harness gives the LLM a master builder's principles and a library of building helpers.
RevenueCat Shipaton 2026 · September 2026
64 / 64 right codes, 0 wrong fills.
SkiPass fills one-time codes from any email inbox into iOS with one tap.
Instalily × Google DeepMind · Toronto 26 · September 2026
#7 · 0.7903 / 1.0
I built forecasters for 10 black-box simulators. Each one predicts 4,000 steps ahead with no feedback.
Experience
Writing
Technical report · Zenodo · Mar 2026 · doi:10.5281/zenodo.19354705
The report treats long-conversation memory as a problem of structuring dialogue on the user's device. Dialogue is compiled into a three-level correlation diagram (sun, planet, satellite nodes) with coordinates and mass. The server model receives the accumulated mass as an additive bias in its attention. HAMIB and Sanma are built on this design.
Technical report · Zenodo · Mar 2026 · doi:10.5281/zenodo.19362604
The report describes a systems architecture that reduces motion-to-photon latency in VR. The headset keeps a local 3D database that integrates Structure Graph Data for geometry with 3D Gaussian Splatting as the optical layer. It predicts upcoming head turns from neck sEMG before the motion is visible and ray-casts candidate future viewpoints against that geometry. The IMU reading then selects the final frame. The headset renders disocclusions and reflections from stored structure instead of guessing them from the current image.
About