Production agentic AI systems in Python: retrieval-grounded assistants, multi-agent crews, and the engineering around them — evaluation, guardrails, cost control and deployment.
Live demo: AskGeorge — an AI representative that answers recruiter questions and assesses fit against a job description you paste in.
| Project | What it is | Run it |
|---|---|---|
| AskGeorge | A deployed AI representative: hybrid retrieval over a Markdown knowledge base, streaming replies, job-fit analysis with a score computed in code so it cannot flatter | uv run python -m askgeorge.app |
| OpenAI Agents SDK | One notebook covering the framework end to end: tools, memory, orchestration by code and by LLM, structured outputs, guardrails | open the notebook |
| CrewAI | Four crews of increasing sophistication, from a two-agent debate to a four-agent team that writes and tests its own code inside Docker | cd crewai/<crew> && uv run <crew> |
Engineering behind the demo: a 23-question retrieval eval that blocks a deploy when recall regresses, an LLM guardrail running in parallel with generation, per-IP and global rate limits, two interchangeable agent backends (framework and from-scratch), and continuous deployment to Modal that scales to zero.
An AI representative for recruiters and hiring managers. It answers questions in the first person as George Traskas, grounded strictly in his real background, and it assesses how well he fits a job description you paste in, with gaps stated plainly. Hybrid dense and sparse retrieval over a Markdown corpus, streaming replies in Gradio, any LLM via OpenRouter, deployed on Modal.
The design decisions and their reasoning are written up in askgeorge/README.md; the build walkthrough is in askgeorge.ipynb (run it from the repo root).
A single notebook covering the framework end to end: agents and the agent loop, tracing, streaming, function tools, memory, orchestration by code and by LLM (agents as tools and handoffs), structured outputs and guardrails. The running example is a small writing team — three agents draft a LinkedIn post in different voices, an editor picks the best one, and a tool saves it.
Multi-agent crews where agents and tasks are declared in YAML, each crew a self-contained uv project.
- debate — a debater argues both sides of a motion and a judge picks the winner on the merits alone.
- financial_researcher — a researcher gathers live information on a company via web search and an analyst turns it into a structured report.
- stock_picker — a manager agent delegates finding, researching and picking a trending company: hierarchical process, structured outputs, a custom notification tool, and memory across runs.
- engineering_team — four agents design, build, demo and test a working system from a paragraph of requirements, executing their own code inside Docker and consulting live documentation over MCP.
Uses uv. From the repo root:
uv syncCopy .env.example to .env and fill in the keys you need. Each CrewAI crew is a
separate project with its own dependencies — see crewai/README.md.
MIT — see LICENSE.