Sideband Academy

How it’s built

Sideband Academy is my exploration of one agentic pipeline that builds a learning system and the content it teaches as a single platform. To a student it’s simple: free, offline and no account. Underneath are a learning engine grounded in learning science, an AI coach that can’t give the answer away and an audio production line that turns the exam pools into narrated stories, all built and maintained by a fleet of AI agents under my direction.

A layered learning engine

Practice runs on SM-2 spaced repetition, adjusted by how confident the student says they are, so a confident wrong answer costs more than a guess. A second scheduler runs mastery campaigns across each pool: a question earns its three marks only through correct, unaided answers spaced well apart, and each session mixes 40 percent weak, 35 percent half-learned and 25 percent new material.

Learning science in every session

Up to one question in four asks the student to recall the answer before the choices appear. Related questions from the same official group test whether learning transfers, confident misconceptions stay flagged until they are repaired, and eight concept labs, on topics such as Ohm’s law and decibels, are recommended when a pattern of errors shows up.

Progress measured by evidence

Readiness has to be earned with evidence: high scores on three timed mock exams taken on separate days, strong accuracy in every section of the exam and one cold exam a day after the last study session. Retention checks return at 7, 14 and 30 days, a daily plan paces study back from the exam date, and rewards go to retention and repaired misconceptions, never to clicks or speed.

An AI coach that can’t spoil the answer

Coaching runs on OpenAI with structured outputs. Before an answer, the tutor offers a single Socratic hint, and the request carries no field for the choices or the answer key, with a deterministic leak detector checking every reply as well. The coach builds study plans from progress kept on the device, and every change runs an evaluation suite of 112 cases plus red-team sets for scope and leaks.

Story Labs, an audio production line

Agents wrote the scripts through a skill built for teaching radio, mapped each story to the exam questions it covers and scored it against a teaching rubric. ElevenLabs voices narrate them, speech recognition checks the audio against the script, and every file gets two-pass EBU R128 loudness mastering. The result is 44 stories, about seven and a half hours tied to 349 Technician questions, with lock-screen and car controls for study on the go.

One platform, one pipeline

A bug report filed in the app becomes a GitHub issue with personal data stripped, labelled for the model best suited to take it on. Every change runs validation against the official question pools, thousands of tests across three browsers and the AI evaluation, and a single command rebuilds the audio catalog the app reads, so the software and its content move together.