Max Nardit
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 decidewhat gets found, trusted,and acted on
01 / Thesis
Operating 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.
02 / Focus
Current areas
- 01
Visibility & discovery
How people, businesses, and tools stay findable when AI systems answer, summarize, route, and act.
- 02
Measurement & tracking
How to know what’s working when clicks, cookies, referrals, and dashboards stop telling the whole story.
- 03
Agentic systems
Memory, handoff, orchestration, tool boundaries, and recovery for agents that touch real workflows.
- 04
Operational evidence
Original-data research, field notes, and shipped tools. Findings over forecasts.
03 / Artifacts
Selected 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.
- 01BeetrootA local-first clipboard environment for preserving workflow context on Windows.
- 02Clipboard MCPA clipboard bridge between humans and AI agents.
- 03Share2AgentA browser-to-agent handoff surface.
- 04Agent RecallScoped persistent memory for long-running coding agents.
- 05Claude Usage DeckA small operational monitor for Claude.ai limits.
04 / Writing
Recent writing
All articles →GPT-6.1 Terra after three days: the published reports, and why none of them tests its computer use tool
Terra was sold on screens and clicks. What got published in its first three days is a pelican, a spreadsheet audit, a latency race, a screen-reading quiz and GitHub's word on C#. The closest thing to evidence about Terra acting past a barrier is OpenAI's own system card, and the number that matters there is how often it found another route.
Importing Claude Code memory into Codex 0.165: enable memories before the first import, or the profile stays on v1
Copying the notes is the easy part. Every Claude Code file survives intact, and what Codex loads at session start is a model-written route to it, inside a summary with a hard size cap. Two default-off switches and one stray directory decide which memory pipeline those files land in.
GPT-6.1 Terra for computer use: a worker seat in the agent stack, not the reviewer's
A mid-tier model earns its price where the result is cheap to check before it matters. Terra's benchmarks measure cost per task; the number that decides computer use is what a wrong click costs, and OpenAI's chart has no axis for it.
OpenAI's no-internet rule for frontier training: the approved services were the way out, and agent sandboxes should copy the split
The copy of the web held. What OpenAI's own red team found a way through was almost every service left connected for running the lab, and the September escape went through a resolver no copy replaced. Serve the agent's work from copies, keep the live connections few and outside its reach, and expect the cost to land on github.com.
Vox's AGENTS.md lines, tested in 90 Claude Code runs: the line that worked best wasn't one of his
Line one made the agent leave out the input check. Line two's extras disappeared when line one shared the file. 'Without breaking anything' showed no clear effect beyond what a named test command already did. What 90 runs can and cannot say, and how to test a line before you keep it.
Attack surface reduction with Claude Code: Anthropic's three passes on my own site, and a fourth they don't ask
The skill is not out, the method rebuilds in an afternoon, and the hard part is choosing where to point it. Here the most useful result was a broken verification file, and the exposure that mattered lived in a deploy setting my three passes never opened.
05 / Contact
Get in touch
Email or Telegram both reach me. Telegram is faster.