Junshuo Lu

An AI learning engine built on spaced repetition and my own Socratic teaching method: the agent asks before it answers, so the understanding stays with the learner. I am its first user, which keeps the product honest.

FastAPI, PostgreSQL, Redis and Docker Compose on a 15-cycle roadmap toward retrieval and agent orchestration. Three cycles shipped: a three-service dev environment with named-volume persistence, hot reload and health checks, then a layered FastAPI skeleton with an async engine and pooled connections. REST design and authentication are in progress. Every milestone is proven with logs and curl, never assumed.

A dependency-free, single-file portfolio drawn as an engineering sheet. A WebGL2 shader renders brushed titanium with anisotropic highlights and heat-tint colour, and your cursor is the only light source: the name breathes toward it on a variable-font width axis, plates cast shadows away from it, buttons lean into it.

Concept, spec and acceptance are mine; the code was generated with Claude Code and verified with jsdom smoke tests and the Khronos GLSL validator. Press / anywhere to open the console.

Five custom skills: a guided learning engine, a reader with six genre models, a note engine for Obsidian, an epistemic guard that refuses to make things up, and a skill creator. Around them, a nine-project taxonomy routes work by identity and mode, and an Obsidian second brain keeps what is worth keeping.

It is the system I think in, so it is also where most of my prompt-architecture judgement comes from.

A web-based music composition tool rebuilt from scratch for a real client, a music educator, with a six-person Scrum team in the UNSW capstone. Score rendering with VexFlow, playback with Tone.js, sprints tracked in Jira, delivered through a branch-and-pull-request workflow. Graded Distinction.

My part: MIDI export on the frontend with @tonejs/midi, a MusicXML import service in FastAPI with music21, and save and load of local JSON project files. Re-reading what the client had actually asked for showed that nobody had requested login or authentication, so the team moved from a full backend to a hybrid: frontend for most exports, FastAPI only where MusicXML parsing needed it.