This week we covered the agentic shift and the push toward efficient, specialized models. Here are three weekend projects that put those trends into your hands.
Project 1: Build a Personal AI Agent with Tool Use
Use Anthropic's Claude API (or OpenAI's function calling) to build a simple agent that can check your calendar, draft emails, and search the web — all from a single text prompt. The key is giving the model access to tools (APIs) and letting it decide which to use.
- Tools: Python, Claude/OpenAI API, Google Calendar API
- Time: 3-4 hours
- What you'll learn: How agentic AI actually works under the hood — tool definitions, chain-of-thought planning, and execution loops
Project 2: Run a Local LLM on Your Own Machine
Download and run a quantized open-source model (like IBM Granite or a small Llama variant) entirely on your laptop using Ollama. No cloud, no API keys, no costs. Then connect it to a simple chat interface.
- Tools: Ollama, any 8GB+ RAM machine, a terminal
- Time: 1-2 hours
- What you'll learn: What "edge AI" actually feels like, model quantization trade-offs, and the surprising quality of small models
Project 3: Automate Your Invoice Processing
Take the document parsing approach Unstructured pioneered (covered in Wednesday's post) and build a simple pipeline that extracts line items, totals, and dates from PDF invoices using an AI model. Feed it 5-10 sample invoices and watch it work.
- Tools: Python, PyMuPDF or pdfplumber, OpenAI or local LLM
- Time: 4-5 hours
- What you'll learn: Real-world document AI, structured data extraction, and why this is a game-changer for small businesses buried in paperwork
Pick one, build it, and you'll understand more about where AI is headed than any amount of reading can teach you. Happy building.