Lucas WintherA notebook of things I've built: what they are, how they work, and who (or what) wrote each part.

What and why

A compute job system for researchers at Aarhus University. Slurm is powerful but unfriendly; this puts a web app in front of it, so submitting, watching and collecting a job doesn't need a terminal.

Both halves scale: the web stack and the compute nodes grow and shrink on Oracle Cloud with the queue, so the university isn't paying for idle machines.

How it works

Who wrote what not marked yet

Part	Written by	Note
React front end	?	Job form, queue view, results.
Node API & job service	?	Talks to Slurm on the user's behalf.
OCI autoscaling	?	Web stack and compute, separately.
OAuth / Entra ID login	?	University accounts only.
ERDA integration	?	Pick inputs, deliver results.

share by hand vs. AI: to be filled in

Decisions placeholder

1. Keep Slurm, wrap it. Researchers' existing scripts keep working.
2. Scale the web stack and the compute pool independently: a busy week of uploads isn't a busy week of jobs.
3. Store nothing long-term in the platform; ERDA is the source of truth.

What broke placeholder

• New nodes joined the cluster before their mounts were ready, and the first jobs on them failed.
• Nodes now report healthy only after a mount check; Slurm waits for that.

What and why

One calendar for the limited-time events in several gacha games, so banners stop slipping past unnoticed. Also a deliberate test: how far can an agent take a small, real project if I only write the spec and the checks?

How it works

Decisions

1. Scrape the announcement pages, not the games.
2. Everything stored in UTC, shown in the visitor's zone.
3. The agent writes code; it never writes the tests' expectations.

Who wrote what detail is placeholder

Part	Written by	Note
Spec & tests	by hand	One page of requirements, one failing test per game.
Scrapers	AI	Written against my fixtures.
Normaliser	AI	First version guessed time zones. Rejected.
Calendar UI	AI	Plain JS, no framework, on request.
Review	by hand	Every change read before merge.

≈ 10% of lines by hand, 90% AI · highlighter = AI

What broke placeholder

• Two scrapers died when a site changed its markup.
• Fixed by the agent in one pass from the failing fixtures.
• Duplicate events after a daylight-saving change.

What and why

Every machine in the homelab, rebuildable from bare metal with one command, with git as the documentation. Deployment management for a growing pile of docker-compose services.

How it works

Who wrote what detail is placeholder

Part	Written by	Note
Inventory & layout	by hand	Which service lives where, and why.
Playbooks & roles	AI	Generated, then iterated for a month.
Compose files	AI	Templated from one pattern I wrote.

≈ 15% by hand

Decisions

1. Compose, not Kubernetes: the lab is three machines.
2. Secrets in an encrypted vault file, never in the repo history.

What broke placeholder

• Every run restarted every container: a template wrote a fresh timestamp on each render.
• Found by hand with --diff. The agent kept suggesting handlers.

What and why

Peer-to-peer archives only work if peers actually keep what they promise to keep. The thesis looks at reputation systems for P2P archival: how peers can rate each other so that free-riders and bad actors are found and routed around.

How it works

Who wrote what

Part	Written by	Note
Simulator	by hand	Node.js, peers as processes in one event loop.
Reputation models	by hand	Compared against a no-reputation baseline.
Report	by hand	Written in 2022, before any of this was AI.

100% by hand

Decisions placeholder

1. Simulate instead of deploy: thousands of peers on one laptop, reproducible runs.
2. Model churn explicitly; honest peers going offline shouldn't look like cheating.

What broke placeholder

• Early runs punished honest peers on slow links harder than free-riders.
• Audits got a grace window, and the result became a chapter.

What and why

A rowing game you play with two real oars. Sensors on the oars feed an Arduino, the Arduino feeds Unity, and the boat in the headset moves the way your arms do. The question: does real resistance make a virtual space feel more real?

How it works

Who wrote what

Part	Written by	Note
Arduino firmware	by hand	Reads two sensors, streams angles.
Unity game & physics	by hand	C#, group project.
Oar rig	by hand	Literally, with tape.

100% by hand

Decisions placeholder

1. Measure oar angle, not force: cheaper sensors, and it's what the boat needs.
2. Smooth on the Unity side so the firmware stays dumb.

What broke placeholder

• Serial lag made the boat feel drunk.
• Smaller packets, higher baud rate, and the latency went away.