TeachBack: You don't know it, until you can teach it back!

Built for a friend learning Python after exhausting shifts at an electricity distribution company.

The Friend & The Friction

My buddy works a grueling 9 to 5 at a local electricity distribution company. He's been trying to break into Python, but he's trapped in the classic "tutorial hell": endless YouTube playlists, passive video watching, and synthetic projects like "build a calculator" that have zero relevance to his actual life.

We kept promising each other we'd jump on Discord for study sessions, but life kept getting in the way. By the time he finishes dealing with feeder voltages, blackout reports, and transformer glitches, he's drained and barely has the energy to look at a keyboard, let alone schedule a live call. Our working hours almost never aligned. I realized he needed something he could run entirely on his own schedule: a self paced learning tool that could act as a live study partner without needing me (or anyone else) on the other end of a call.

The Spark & The Idea

Around the same time, I applied to teach for Stanford's Code in Place. As part of the application, they had me teach a simulated classroom of three virtual students who watched my screen, listened, and asked sharp questions.

It clicked: the fastest way to figure out if you actually understand something isn't reading about it, it's explaining it to someone else.

Most platforms follow a rigid pipeline:
Theory
⬇️
Practice
⬇️
Workshop.
But passive comprehension is a trap; you think you get it until you try to say it out loud. I wanted to add the crucial missing fourth step: "Teach."

What I Built

TeachBack is a lightweight learning platform you use in the browser, run from a small local server on your own computer. It's tailored to my friend's reality:

8 tailored lessons using data from his world: meter readings, line loads, and outage priorities.

• In-browser Python, so there's no Python installation to wrestle with after an 8-hour shift.
• The "Teach" phase: instead of a quiz, he explains the code to three distinct AI students:
  • Maya: the curious one who always asks why.
  • Kofi: the edge-case guy who asks what breaks when inputs go sideways.
  • Zee: the beginner who needs everything in plain, simple words.
• An objective examiner: a second AI agent that reads what he says (as text) and checks whether he explained each core idea in his own words. If it can't point to words he actually said, the explanation is rejected.
• The Capstone: his final project is to design his own certificate in Python. Once he builds it and teaches the code back, the app shows the certificate he made.

AliFikan94 / hacktober26

🎓 TeachBack

Honest status (hackathon submission). TeachBack's lessons, in-browser Python runner, capstone, certificate, model picker, privacy panel and the Teach logic (students, examiner, voice loop) are implemented and covered by automated tests and browser tests against simulated model servers. On the author's low-memory PC (about 2.5 GB free RAM) the Teach conversation was not verified end to end with a real local Gemma before the deadline, and its cause was not pinned down. The in-app Test the students and Test the microphone buttons (model menu, top right) and npm run doctor exist to diagnose exactly that. It has not been tested by the friend it was built for.

Theory → Practice → Workshop → Teach. A learn-and-teach platform built for a friend who works 9–5 at an electricity distribution company, wants to learn Python, and is too tired to type after work.

Each lesson has four phases: 📖 Theory…

View on GitHub

Repository: github.com/AliFikan94/hacktober26

How I Built It

Open-source AI and tools

• Open-weight models through Ollama running locally. My target was Gemma 3 (I pulled gemma3:4b and gemma3:1b), reached through Ollama's OpenAI-compatible endpoint.
• Mastra for the agents: four of them (Maya, Kofi, Zee and the examiner), each with its own instructions. A simple rule picks which student speaks, based on what the learner just said.
• Pyodide (Python compiled to WebAssembly) so Python runs inside the page.
• A small TypeScript/Node server streams each student's reply to the browser as soon as it's ready. The browser's built-in speech tools handle voice.

How the examiner stays honest: it must quote the learner's own words as evidence, and the server rejects any quote that isn't really in what they said. If a small model can't answer in the required format, a labelled keyword check takes over.

Built with AI assistance: I built this with an AI coding assistant (Claude Code), working under my direction.

The Honest Part: What Didn't Work

I'm submitting this with complete honesty about its rough edges:

• Hardware constraints: A 4B Gemma model needs around 4 GB of free RAM, but my test PC only had about 2.5 GB. I built an automatic memory check that steps down to a smaller model, or to a cloud model (and says so), when hardware is tight.
• Small model formatting: Tiny local models love to answer in paragraphs instead of structured JSON. Hence the fallback keyword check.
• Voice fragility: Browser speech recognition behaves differently across platforms. I added microphone diagnostics with plain-English error messages.
• End-to-end testing: The Teach system passes automated tests against simulated models, but I couldn't get the full local conversation working on my low-spec machine before the deadline. My friend also hasn't done a full run-through yet, so I don't have a quote from him... yet!

What I Learned

Building an AI tool for a real person with a real job and real hardware limits is humbling. The hard part wasn't writing prompts, it was making open-source AI light enough to run on an everyday laptop. Next: a lightweight hosted option for low-spec hardware, and getting my friend into a live session to test his outage-priority scripts.

Why Does Open Innovation Matter?

For a tired adult learner who is learning in the evenings, open models matter in three ways:

1. Privacy: his half-baked logic and "dumb questions" can stay on his own machine. (Browser voice recognition in Chrome and Edge sends audio to external servers, so the app tells you plainly that typing is the fully private option.)
2. Cost: a local model costs nothing to run every evening.
3. Control: I can tune the students' personalities and the examiner's rules by editing prompts, which gives me far more control than a closed, rigid chatbot. I could also swap the model with one setting, which is exactly what let me adapt when my PC couldn't run the first one.

Prize Categories

• Best Use of Mastra: four Mastra agents drive the students and the examiner.

(Gemma was my target model, but I couldn't show it running end to end, so I'm not entering that category.)