Skip to content
Projects

Independent

Bingo

An AI learning workspace in design: your goal and sources become a study plan, a personal tutor, and realistic practice, using the models you choose.

Role
Software Engineer
When
2026 – now
  • Any

    AI model you choose, including self-hosted

  • 3

    devices from one web app: desktop, iPad, and iPhone

  • 4

    delivery phases, from tutor to full Studio

  1. Situation

    Learners have limited time, may not know where to begin or where to find credible sources, and struggle to see whether they understand more than before.

  2. Task

    Design Bingo, an AI learning workspace that turns a goal and the learner’s sources into a study plan, a personal tutor, and realistic practice using the models they choose, as a system a small team can build and run.

  3. Action

    • Organized the product around notebooks, one per class, exam, or topic, each with Sources, a Tutor, and a Studio of study tools. Members share materials while each learner’s progress stays their own.
    • Chose a Python modular monolith: Django and Django Ninja own accounts, permissions, learning rules, grading, and spending, behind a React and Next.js web app that installs as a PWA on desktop, iPad, and iPhone.
    • Designed retrieval on PostgreSQL with pgvector and native full-text search, so the tutor teaches from the learner’s own sources.
    • Put the tutor’s workflows on LangGraph, with Django services owning the rules, and every model call behind one provider interface built on LiteLLM, so learners can use any model they like, hosted or self-hosted, with Manual and Auto routing.
    • Made long work durable: conversations and job state live in PostgreSQL, Celery and Valkey run the jobs, and client-generated IDs keep a reconnect from starting a second billable generation.
    • Tracked spending in a PostgreSQL ledger that reserves a maximum before each job and settles actual usage, and kept saved API keys encrypted and out of prompts, logs, and the browser.
    • Planned one global AWS deployment with Docker Compose, Caddy, and OpenTelemetry, and four delivery phases with release checks, such as no data crossing between notebooks and exams that resume after a worker fails.
  4. Result

    • The product and architecture spec is at revision 9, with an engineering blueprint for the first build.
    • The build starts one segment at a time, beginning with the local environment.
    • Features are planned, not shipped: each one counts as implemented only when working code passes its checks.

Stack

Python / Django / Django Ninja / Next.js / TypeScript / PostgreSQL / pgvector / LangGraph / LiteLLM / Celery / Valkey / Docker Compose / Caddy

Engineer resume