AI engineering projects are easier to find than to finish. If you are learning AI engineering and stuck on what to build, Landed's GitHub catalog, ai-engineer-portfolio-projects, lists 85 project ideas from beginner to advanced across AI agents, RAG systems, fine-tuning and even language models built from scratch. Each idea lists the skills you will practise, its difficulty and links to help you learn. The reel below walks through the catalog, and this post adds the part the catalog does not: how to choose one project at your level, build a working demo and show what you learned. The presenter in the reel is an AI avatar of Hammad; the words and facts are his.
What is in Landed's AI engineering projects catalog
The catalog is maintained by Landed, a job-search company, and the content is MIT licensed, while every referenced repo keeps its own licence. When I checked it on 20 October 2026 it listed 85 projects in 10 themes: 12 RAG apps, 12 agents and tool-use projects, 10 evals and LLMOps projects, 8 fine-tuning and training projects, 8 multimodal projects, 7 structured extraction projects, 7 LLM from-scratch and internals projects, 8 production and serving projects, 6 prompt and DSPy projects and 7 GTM and AI-PM prototypes. Each theme has its own markdown file with the full list, a one-line description of what you build, the skills it proves, a difficulty rating of beginner, intermediate or advanced, and one to three typed, licence-noted reference links.
They are project briefs, not tutorials
This is the caveat I put in the reel and on the resource page. An entry in the catalog is a brief plus references: a paper, a docs page, a course or a repo. It is not always a complete tutorial you can follow end to end, and the linked tools carry their own licences and hosting or model costs. The README is also upfront that one of the hubs it recommends, RAG_Techniques, is under a non-commercial licence, so you should re-implement its techniques rather than copy its notebooks into a repo you publish. Read the licence line next to every reference before you build on it.
A beginner RAG project to start with: Document Q&A
In the reel I open the RAG category first because the catalog calls it the highest-signal theme. The first entry, Document Q&A RAG, is rated beginner. You ingest a corpus you own, such as your own PDFs, a Slack export or a Notion workspace, chunk and embed it, retrieve the relevant passages and answer with inline citations. The skills it lists are chunking strategy, embeddings, top-k retrieval, prompt grounding and citations, and the catalog suggests making retrieval hybrid, BM25 plus dense, to stand out from the tutorial crowd. The rest of the RAG file steps up from there: contextual chunking with a measured recall lift, query rewriting, agentic RAG, GraphRAG, corrective RAG with a hallucination checker, text-to-SQL and multi-modal RAG over page images.
How to scope an AI engineering project you can finish
ONE TACTIC A WEEK
My checklist from the resource page is six steps. Choose Document Q&A and use a few documents you own or have permission to use. Define one question your demo should answer and one it should decline when the answer is absent. Build the smallest version: ingest, retrieve, answer with sources. Test at least ten questions, including missing answers and conflicting documents, and record what fails. Publish a short demo with setup instructions, sample inputs and results, with secrets and private documents removed. Then explain one tradeoff, whether accuracy, speed, cost or simplicity. A specific limitation is more useful than an unsupported claim.
A planning prompt for Claude or any coding assistant
The resource page has a copyable prompt that does the scoping for you. You give it the catalog URL, your current skills, the hours you have this week and your maximum spend, and ask for one suitable idea with the reason, a smallest working version, milestones, ten evaluation cases, expected dependencies and costs, and a README outline. The two lines that matter most are at the end: separate verified facts from assumptions, and do not invent integrations, benchmark results or work I have not done. Those lines stop the assistant writing a portfolio README for a project that does not exist yet.
What the catalog says a reviewer looks for
The README's own advice is depth over breadth: pick the theme where your portfolio is thinnest, ship a few projects with real evals and a live URL, and write each README as a product spec with the problem, an architecture diagram, eval numbers and cost. Its red flags are a GPT-4 wrapper with no original work, no evals, benchmark claims without numbers and five copy-pasted Streamlit demos. That matches my own rule before you share a demo: it should run from the README on a clean environment, show an example input and output within a minute, document failures, costs and required keys, back every citation with the source text, and contain no tokens, customer data or personal records. None of this guarantees interviews or a job, and I would not claim it does.
Get the catalog and build checklist
The free guide at withhammad.com/resources/ai-engineering-projects links the catalog, walks through the first build, includes the planning prompt and ends with the checklist to run before you share a demo. On Instagram, comment PROJECTS on the reel and the guide link arrives by DM (follow @withhammad.ai first so the DM reaches you).
Related reading
If this is useful, see how to run a job search with Claude and the AI website design brief for your AI coding assistant.