AI/ML

AgentBricks

Open-source ML learning platform — build production-grade systems inside a synthetic agent universe

Personal

6 technologies
2 key decisions
4 results

Problem

Problem

ML education defaults to toy datasets, tutorial code, and isolated concepts that don't translate to production. The gap between "I followed the tutorial" and "I can build a real ML system" is enormous. AgentBricks addresses this by giving learners production-grade infrastructure problems — Kafka streams, real databases, cloud deployment — but fueled by millions of synthetic agents rather than real user data, eliminating privacy concerns while preserving scale.

Approach

Approach

AgentBricks is structured as story-driven modules where each arc presents a realistic business scenario: a synthetic universe of millions of agents behaves like real users — buying, churning, clicking, and converting — and the learner must build the ML system to respond. Modules ship with real infrastructure: Kafka for event streaming, PostgreSQL for storage, Docker Compose for the full stack. Each completed module produces a showcase-ready portfolio project, not a notebook.

Architecture

Architecture

AgentBricks — system diagram

AgentBricks CLIStory ModuleSynthetic Agent Ge…Kafka Event StreamML PipelinePostgreSQL

Key Technical Decisions

Key Technical Decisions

Assembly Instructions — 2 Steps
01

Story-driven framing over tutorial problems

Toy problems don't motivate production thinking. When a learner is debugging a recommendation system for 'millions of users generating 50K events/second', they make fundamentally different architectural decisions than when working on the Iris dataset. Narrative context changes what good looks like.

02

Real infrastructure from day one

Most ML courses use mocked or simulated infrastructure. AgentBricks ships with actual Kafka, actual PostgreSQL, and actual Docker Compose. The learning curve is steeper, but the outcome is a real system the learner can talk through in a job interview.

Results

Results

  • Production-grade ML curriculum using real Kafka, PostgreSQL, and Docker infrastructure
  • Synthetic agent generator produces millions of realistic behavioral events
  • Each completed module generates a portfolio-ready project, not a notebook
  • Zero privacy concerns — no real user data involved

Tech Stack

Tech Stack

PythonDockerKafkaPostgreSQLpytestGitHub Actions

Links