hey, I'm jasper.

Canvas was about to take my course files away at the end of a semester, so I wrote an extension that grabs all of it: files, pages, assignments, discussions, grades, quizzes, even lecture videos with captions. It exports to offline HTML or Markdown, which turns out to be a great way to feed a course into NotebookLM or Claude.

I put it on the Chrome Web Store and forgot about it. Then the Canvas data breach in May sent a wave of students and teachers looking for backups, and it was already sitting there. Strangers started sending pull requests. It has a Featured badge now, 4.8 stars and 30 releases.

Point it at an article, a Canvas page, a PDF or a video and it makes teaching material: summaries, quiz banks, flashcards, rubrics, slide outlines. Exports to Word, PowerPoint in the VU template, Canvas, Moodle and Anki. On a Canvas course it reads the whole course first, so the quiz matches what was actually taught.

It runs on the VU’s EU Azure setup and never stores page text. I wrote the threat model too, since a university privacy review is its own boss fight.

All 35 courses of the bachelor, 515+ quiz questions, 176 practice problems, summaries, grade calculators and an AI chat for each course. The material was spread over six platforms, so I put it in one place.

875 people found it in the first two weeks without me posting it anywhere. Exam week does about 200 a day.

Every semester I talk to students from each year, take their feedback and update it. It’s never really done.

An MCP server for Marktplaats and 2dehands. Your assistant can search, check sellers and compare prices, and with your own login it can message and bid too. There’s a hosted read-only endpoint you can paste straight into claude.ai.

It’s in the official MCP registry, has install guides for 40 clients and a CI canary that complains when Marktplaats changes their API.

Speech goes to text, a BERT model picks the intent and pulls out the slots (cuisine, ingredients, diet, time, exclusions), a MARBEL dialogue manager keeps the conversation on track and a Prolog database narrows down the recipes as you talk. Fourteen intents, with recovery paths for unclear requests and empty results.

The saved evaluation run scores 94.7% intent accuracy on 2,772 test utterances. Built with Arjun, Omar, Kyan and Skick for the Project Conversational Agents course at VU.

Trip planning is still eight tabs and a calculator. So at bunq’s hackathon we built an agent that lives inside the bank. You type or talk, it asks one question, does its web research where you can see it, and pitches three packages. Say yes and it fires seven real sandbox calls: opens a savings pot, funds it, books the hotel through an actual browser, pays, pings your phone for approval, sets up weekly savings and splits the bill.

What I took home: build the demo path first, make the work visible, and boring beats flashy. We swapped headless search for plain HTML search late at night. It never crashed on stage.

Suppliers, menus, prices, staff, marketing, all decided by the agent. We tried an LLM in the live loop. It scored worse than plain rules, so we cut it and had Claude suggest rule changes between runs instead.

Every change needed a mechanism we could name and a signal we could see. The final eval ran on seeds nobody had seen before: 20 out of 20 restaurants survived, zero bankruptcies. Teams with higher peaks went bust.

An open-source MCP server with 20+ tools and a Common Cartridge toolkit. Give Claude Code a syllabus and it builds modules, pages, assignments and quizzes in one import. Everything stays unpublished until the teacher has looked at it, because nobody wants a robot publishing their course at 3am.

A native API client with zero dependencies that idles at about 30 MB of RAM. No account, no telemetry. It imports Postman collections, stores requests as plain files and keeps secrets in the Keychain. I measured Postman at 353 MB and got annoyed. Install it from my own Homebrew tap.

A faithful simulator of Beverbende, a PPO agent that learns by playing itself and beats every scripted bot over 40,800 rounds, and a short paper written for people new to ML. One “style” input changes when it knocks, from turn 14 to turn 2.

Then a Svelte game where you play the trained network, with a custom otter deck. The TypeScript engine is tested move for move against the Python one.

Rooms in Amsterdam are gone in minutes, so this watches Kamernet, scores every new listing with an LLM against what you’re looking for and pings you wherever you are: Discord, Telegram, ntfy, email, WhatsApp. Self-hosted, Docker, optional dashboard. Also available as an MCP server so an agent can search for you.

Get a Discord, Telegram or ntfy message the moment a saved search matches or a listing drops in price. Self-hosted in Docker. It started because I wanted a secondhand Apple TV 4K and refused to keep refreshing.

Two reinforcement learning methods, the same level, the same compute budget. We compared learning speed, stability and final distance for A3C and PPO over four random seeds and up to 3 million timesteps each, and wrote it up as a paper. Both reached the flag. PPO got further on average, A3C was noisier.