Alex, a learning assistant for Slack
Alex turns everyday work questions in Slack into short, personalised learning paths. I built it alone: a Slack bot on a LangGraph state machine, with a FastAPI backend, Supabase and deployment on Google Cloud Run.
- When
- around Jan to Feb 2026
- Status
- Offline The Cloud Run service is switched off.
- Where
- Own project
- Built with
- LangGraph, Slack Bolt, FastAPI, Supabase, GCP
- Made as
- Individual project
- Goes with
- AI micro-course learning platform
About the project
People ask Alex what they want to learn, right in Slack. After two short questions (why, and how experienced they are), Alex proposes a learning path of three milestones, each with a starting point and a goal. The learner can edit a milestone, regenerate the path or start over before accepting it.
Alex then teaches each milestone in short chunks with a progress bar, asks the learner to reflect in their own words, and closes with a quick check. Every answer gets an explanation, and the path moves on to the next milestone.
Under the hood, a LangGraph state machine routes each message through guardrails, onboarding, intent detection, content generation, fact grounding with Perplexity (Gemini as backup) and a check that the content matches the learning goal, regenerating when it does not. Lessons can also be read aloud with Gemini text-to-speech, falling back to Google Cloud text-to-speech, and learners can answer by voice.
Sessions and conversation history live in Supabase (PostgreSQL), runs are traced in LangSmith, and a FastAPI service reports health and metrics. The bot is packaged with Docker; it ran locally for the demo and was deployed to Google Cloud Run for testing, not kept online. The repository has a suite of about 245 automated tests.
Double-tap the left or right side to jump 10 seconds back or forward, or use the arrow keys.
How it is built
Alex runs as a single container. The Slack connection uses Socket Mode, an outgoing WebSocket from the bot to Slack, so nothing has to be open to the internet; a small FastAPI service next to it answers health checks and reports usage per organisation.
Every message goes through a LangGraph state machine that checks it against guardrails, knows where the learner is (onboarding, path, milestone, lesson) and picks the model for each job. The session lives in Supabase and is written in one atomic upsert, because two Slack events for the same person can arrive at the same moment.