白石 拓海
Founder and engineer at Wovol. Local-first AI products, evidence-linked developer tooling, and the unglamorous safety work that makes them shippable.
Evidence you can actually check. Every claim should link back to where it came from: a source line, a preregistered test, the iteration that failed.
Most of this work lives in private repositories. The links point to what is genuinely public: demos, landing pages and write-ups.
An anticipatory workspace layer for macOS. Press a hotkey and the thing you were about to resume is already prepared.
Swift 6 and SwiftUI, around 240 sources, notarized builds. On-device only: nothing leaves the machine.
One personalized disaster-response agent per citizen, 8,000 of them reasoning in parallel on a single AMD Instinct MI300X.
Cascades are tracked as a chain, not as isolated events. Built for the AMD Developer Cloud Hackathon 2026, Track 3.
0 agents, one MI300XSchematic: one dot per agent
Legacy Java/Spring comprehension for Japanese SIer firms. Every claim in the report links back to the source line that justifies it.
Self-hosted. Emits HTML/PDF reports plus a machine-readable evidence.json, with no managed backend: it runs entirely on the customer's machine.
Does the way code is represented change how frontier LLMs reason about it? An information-parity control (C1+) tests that at constant information.
Preregistered, nine-cell confirmatory family, Holm-corrected. Non-rejection on all nine cells, reported as a null.
Nine preregistered tests: three hypotheses across three tasks.
A mechanical calculator built from marble logic gates: XOR, AND, OR and a custom Y-split, laser-cut from 2D MDF.
A marble present is 1, absent is 0. The repository ships a truth-table evidence pack and the iterations that failed.
| A | B | AND | OR | XOR |
|---|---|---|---|---|
| A 0 | B 0 | AND 0 | OR 0 | XOR 0 |
| A 0 | B 1 | AND 0 | OR 1 | XOR 1 |
| A 1 | B 0 | AND 0 | OR 1 | XOR 1 |
| A 1 | B 1 | AND 1 | OR 1 | XOR 0 |
23°
build angle
23°: 10 trials
Private, in the Wovol monorepo
A joy-first, privacy-first AI study coach. No third-party telemetry by default.
Public
A working SEO method for people who have commit access to the site. Prose, checklists and agent playbooks. No crawler, no scoring engine.
Private
A local-first Claude Code ops lab: a memory layer inspired by neuroscience (mnemos), four-profile browser isolation, and a bilingual safety net that scans every write against 30+ PII and secret patterns. It never auto-sends.
91% additions, across 9 files
Proposed a read-only, least-privilege OAuth scope profile for the tesla_fleet integration.
Closed after discussion with the integration's code owner, who pointed out that the existing consent flow already lets users deselect command scopes. The review thread was worth more than a merge would have been.
PRs merged into production
August to September 2026
Private client repository
Repo-wide secret scanning, git history included. CSP hashes generated in the build. A dedicated security CI workflow. PII matching that covers Western identifiers and Japanese ones: マイナンバー, 保険証 and 戸籍.
The top algorithmic tier on paiza's coding-skill check, in Python.
Japanese native, fluent in English. Product copy, security tooling and customer documents all ship in both.
Most commits land in private repositories, so public language statistics understate the actual mix.
Open to conversations about local-first AI, Apple-platform work, developer tooling for legacy systems, and anything involving evidence you can actually check.