Data & AI Systems Engineer working on visibility, measurement, and agentic systems.
AI now sits between people and what they’re looking for, and between work and the software that runs it. It answers, routes, remembers, forgets, and now acts on its own. Usually before anyone can check whether it got things right. I study that layer with original data and build the tools to work inside it. My background is the supply side of digital marketing: the crawl, the analytics, the automation and reporting, the plumbing under what looked like marketing. The marketing was never really the problem. Visibility, measurement, and control were. AI didn’t make that problem smaller. It moved it somewhere you can’t see.
Beetroot featured in 窓の杜 (Japan) · original-data research · shipped open-source tools
the systems that decide
what gets found, trusted,
and acted on
Thesis
AI is changing two things at once: how people find information, and how work moves through software.
That shift isn’t only a search problem. It touches tracking, attribution, context, memory, handoff, and trust: the systems that decide what a person sees before they make a choice.
I work on that layer. Some of it is research: measuring what changes and publishing what holds up. Some of it is engineering: building tools and workflows that keep context, expose failures, and make AI-assisted work inspectable.
Focus
Writing
How to route a Claude Code subagent to Haiku 5.5 on purpose, which effort level to pin it to, how to keep it under the long-prompt rate with /autocompact, and what the system card says to check before you trust it with a tool loop.
Identical price cards, two different cliffs. On long prompts the models part ways: one reprices the whole request above 100,000 tokens, the other above 272,000. For a subagent that reads files, that line decides the bill more than the rate does.
Half of the ten fixes land on guarantees the documentation makes, not on the limits it already disclaims. Two more are parts that failed open, and for a PreToolUse hook script that crashes or hangs, failing open is still the documented behavior.
Seven places vendors want a typed answer to replace a free-text judge, what Red Hat's benchmark found, and why every vendor in the category, including the one that sells calibration, ends its documentation by handing the threshold back to you.
What a text watermark result can establish, how much editing it survives, and why the error rates you can read so far were all measured by the vendor that holds the key. Then what that means for anyone who grades, writes or builds on these models.
Anthropic's own safety page says it in one line: isolation limits where Claude's code runs, not what Claude reads or does. The cloud move relocates the first and leaves the second wired to your desktop app, which now answers for a session that keeps going after you stop watching.
Artifacts
Tools I’ve built around the same layer: memory, clipboard, agent handoff, monitoring. Local-first where it makes sense, from problems I hit directly.
A local-first clipboard environment for preserving workflow context on Windows. AI transforms, OCR, smart search, unlimited history. Built around the idea that the clipboard is not temporary — it is part of workflow memory.
Learn more →
A clipboard bridge between humans and AI agents. Read, write, and watch clipboard state through MCP. A small interaction primitive for agent-native workflows.
Learn more →
A browser-to-agent handoff surface. Send pages into agent workflows through a webhook, preserving context at the moment of browsing.
Learn more →
Scoped persistent memory for long-running coding agents. Local-first, MCP-native, designed so context survives across sessions without piling up.
Learn more →
A small operational monitor for Claude.ai limits.
Learn more →