I build software, then I make it faster.

AI chat collapses learning into one scrolling column. A useful answer might come from one line of a PDF, lead to three follow-up questions, become a flashcard and later a mock-exam topic, and a transcript loses all of that structure. StudyCanvas keeps the source and the reasoning on the same infinite canvas.

A highlight sends Claude Haiku 5.5 the exact selection, its page context and the parent answer, and the response streams into a linked node. Gemini still handles voice, images, web search and scanned pages, and takes over for text if Claude is ever unavailable. Every in-flight request owns an AbortController, so closing a node cancels the work and settles what has arrived into something readable. Python runs in the browser through Pyodide, so the server never executes a student’s code. Workspaces are local-first: a real folder through the File System Access API, with IndexedDB where that isn’t available.

Every answer, quiz, flashcard and summary carries a confidence slider from Lost to Perfect, and the rating fades over time to decide what comes back first. Blurt tests recall with the book closed: you say what you remember, it maps your words onto the lecture’s key ideas, and it marks what you missed.

Ten students use it every week, including the students I tutor.

You learn a topic from someone in town, then walk into the outskirts where creatures quiz you on it. The four answer buttons are the four move slots, so the quiz is the battle. Every question carries a required explanation, so the teaching moment appears whether you get it right or wrong.

Answers feed a five-box Leitner scheduler on 1, 3, 7, 14 and 30 day intervals. A wrong answer drops a question by one box, so a single slip never wipes out a month of progress. A correct answer only promotes a question that was actually due, so “mastered” has to be earned across weeks of reviews.

We were a team of 6, selected from 1,300+ applicants, and I owned the rendering pipeline. The town is a custom 2D tile renderer. Moving ambient updates across the 792-tile world out of React’s render cycle and into CSS made tile updates 2.5 times faster, and end-to-end frame rate rose 53% once camera, movement and input scheduling were rebalanced.

We are taking it to Web Summit Lisbon in November as an ALPHA startup.

The glasses stream gyroscope, accelerometer and magnetometer readings at 476 Hz over raw HID. I decode them, calibrate the sensor-to-head orientation and estimate pose with a Madgwick fusion filter, written from scratch, that adjusts gyro bias only while the head is at rest. A separate magnetometer calibration enables a slow heading lock: it corrects yaw drift against the local magnetic field and rejects sudden disturbances. Pitch and roll still come from inertial fusion. Together they remove about 20° of yaw drift over a 22-minute worn session and cut on-screen text jitter 6×. Every change to the filter is checked against recorded on-head sessions and 19 synthetic scenarios. A Direct3D 11 renderer places each screen at its own yaw, pitch, roll and distance. The default is a three-screen arc at −45°, 0° and +45°.

The screens are real Windows desktops. Virtual monitors from the Parsec display driver are captured with DXGI Desktop Duplication, with a GDI fallback, so any app you drag onto them works. A native dashboard edits layouts, recentres the view, controls a soft reading hold, per-screen dimming and night tint, and shows device and engine diagnostics.

The first model the extension shipped performed at chance. Its replacement looked better until I caught it learning a dataset shortcut. So I rebuilt the evaluation around content-matched pairs, where each real image has a generated twin of the same subject, before trusting any number.

The shipped detector is a fine-tuned DINOv2-S/14, quantised to INT8 and run through ONNX Runtime Web with WebAssembly. It scores 0.954 AUROC across four diffusion models, holds between 0.948 and 0.958 under JPEG compression, resizing, WebP conversion and screenshot recapture, and abstains when it is unsure. Ten interactive forensic tools sit behind every verdict so you can check it yourself.