AI micro-course learning platform

Generates a personalised short course and an AI tutor for each learner. I built it alone, end to end: LangGraph orchestration, LangSmith tracing, FastAPI services, a PostgreSQL database, the front-end and deployment on Google Cloud.

When
Mar 2025 to Mar 2026 (about a year)
Status
Online Running; sign-ups are closed.
Where
Own project
Built with
LangGraph, LangChain, LangSmith, FastAPI, PostgreSQL, GCP
Made as
Individual project

About the project

The platform builds a short course around what one learner wants to know, and gives them an AI tutor to work through it with.

I delivered the whole solution myself: the AI steps as a deterministic LangGraph workflow, with LangChain for the model calls and tools inside each step, observability with LangSmith so every run can be traced, the database integration, the front-end, and the deployment on Google Cloud Platform.

It started small: in March 2025 the web app was built and hosted with Lovable, and the course generation ran as Make.com scenarios called from it, a quick way to try the idea before building my own services. In August 2025 I moved it to my own FastAPI services on Google Cloud Run, which gave me control over the AI steps, their cost and their tracing.

Double-tap the left or right side to jump 10 seconds back or forward, or use the arrow keys.

A learner asks for a course on leadership for first-time managers. Alex, the AI tutor, runs a short onboarding, the course is generated (sped up), and the learner reads and listens to the slides, asks Alex a question, answers the practice scenarios in all three modules and earns the certificate. Recorded on the live platform. Music: “Inspired” by Kevin MacLeod, licensed under CC BY 4.0.

How it is built

How the AI micro-course platform is builtThe web app talks to Supabase for sign-in, data and files, to Supabase edge functions that start a course, to the AI tutor, and to Mixpanel for product analytics. Two FastAPI services on Google Cloud Run do the AI work with OpenAI, Gemini, Perplexity and Firecrawl; PDFs that learners attach go straight from the browser to Google Cloud Storage.Learner's browserProduct analyticsSupabaseGoogle CloudAI providersWeb appReact, hostedon VercelMixpanelhow learnersuse the appAuthsign-in and usertokensDatabasePostgreSQL; learnerssee only their ownrowsStorageslide images andprofile picturesEdge functionscheck the token,then call a serviceCourse generatorFastAPI on Cloud RunCloud Storagethe PDFs learnersattachAI tutorFastAPI on Cloud RunPerplexity,Firecrawlresearch the topicOpenAIGPT-4o and imagegenerationGeminislide images andnarration121532645687899374162354

The learner's browser only talks to Supabase, a few Supabase edge functions and the AI tutor, plus Mixpanel, which records how learners move through the app (the dashed line). Row-level security in PostgreSQL makes every query return only the learner's own rows, so the web app reads courses straight from the database, with no server in between.

Anything that costs money or needs secret keys sits behind a check. Edge functions verify the learner's token before they start a course and pass a server-only key to the generator; the tutor verifies the token itself and checks that the learner owns the course. The generator is a LangGraph workflow that researches, writes, draws and narrates, then writes the result back to Supabase; every run is traced in LangSmith.

A learner can also build a course from their own PDF. The browser uploads it straight to Google Cloud Storage through a signed link that lasts 15 minutes, so large files never pass through a server; the edge function only accepts files from that upload area, and Gemini reads the PDF directly, tables and layout included, to write the course.

Get in touch

Let’s talk if you have a project in mind.