Build a working AI agent on your laptop — no cloud account, no GPU, no vendor lock-in.
This is a brev-free rebuild of the kind of "build an agent" workshop that normally runs on a managed cloud-GPU platform. Everything here runs on a normal laptop against a free hosted API. Three short parts take you from "what even is an agent" to a retrieval-grounded shopping assistant whose answers you can actually measure.
Built and maintained by MEMOTECH — IT expertise since 1998 · Swiss Based.
An AI agent is software that, given a goal, decides which steps to take instead of following a fixed script. A shopping agent asked for "a warm jacket under CHF 250" decides on its own to search the catalogue, read a product's details, and compose an answer.
This repository teaches that idea by building one — small, in the open, with nothing hidden. By the end you will understand, concretely, what the "AI assistant" in every vendor pitch this year actually is.
The honest part: this tutorial builds a demo. A demo agent and a production agent that touches your real catalogue, your real customers and your real revenue are different animals. The table near the bottom — What this tutorial shows vs. what production needs — is the most important thing on this page if you are evaluating agents for your business.
Three parts. Each is a self-contained folder with its own README:
| Part | You build | Key idea | GPU? |
|---|---|---|---|
| Part 1 — Build an Agent | A tool-calling shopping assistant | An agent is a loop: model → tool → model | No |
| Part 2 — Agentic RAG | The same agent, grounded in a product catalogue | Retrieval, so it answers from real data | No |
| Part 3 — Evaluation & Guardrails | An evaluation + guardrail harness | "Is it good? Is it safe?" — measured, not hoped | No |
Part 1 is built twice on purpose: once as a plain Python loop (so you see what an agent really is), then once with LangGraph (so you see what a framework adds). Nothing in the three parts needs a GPU — it all runs against a hosted API.
Each part also has a companion deep-dive on m3mo Bytes:
- Part 1 — An AI Agent Is Just a Loop
- Part 2 — The Product Your Search Bar Can't Find
- Part 3 — A Fluent Answer and a Correct Answer Look the Same
Full walkthrough in setup.md. Short version:
- Get a free API key at https://build.nvidia.com (no GPU, no credit card).
cp .env.example .envand paste your key.uv sync(orpip install -e .), thenjupyter lab.
The default configuration uses the hosted NVIDIA Nemotron™ model: free
tier, no GPU. But no tutorial code is provider-specific. Every model
call goes through llm.py, which speaks the OpenAI
chat-completions protocol. Point .env at OpenAI, a local
Ollama, or a self-hosted vLLM and every notebook
still runs unchanged. Removing the cloud-platform lock-in is the whole
reason this rebuild exists.
A demo proves the idea. It does not make the idea production-ready. The gap between the two columns below is real engineering work:
| This tutorial (demo) | Your production agent needs |
|---|---|
| Sample catalogue, 12 products | Real catalogue, 10k–100k SKUs, data-quality cleanup |
| A notebook on your laptop | Deployment, scaling, a latency and cost budget per query |
| No guardrails | An agent that never invents prices, delivery dates or product facts |
| A ~25-item evaluation set | Continuous evaluation in production, regression monitoring |
| No data-protection context | Swiss nDSG compliance, data residency |
| A standalone agent | Integration into your real Shopware checkout / Store API |
That right-hand column is what MEMOTECH does. If you want an agent like this running on your product catalogue, start with a free needs analysis: memotech.ch/agentic-commerce · mehmetgoekce@memotech.ch.
Independent clean-room rebuild — inspired by NVIDIA's "build an agent" workshop, but written from scratch with its own code and examples. Not affiliated with, sponsored by, or endorsed by NVIDIA; the tutorial simply uses NVIDIA's publicly available free API as one (swappable) option.
Licensed under Apache-2.0. © 2026 MEMOTECH.