Projects
Each one shipped with a write-up. The status tag says where it stands today.
My Storyland live
Give it a book, get a trip. A multi-agent system reads the novel, finds the places behind it, and builds a walkable itinerary with the book-to-place connection explained at every stop.
Started in November 2025 as a capstone for Google’s five-day AI Agents Intensive on Kaggle, where it placed third in the freestyle track among thousands of entries. Now a live app, with the company mostly run by a team of AI agents. Launching on Uneed Oct 24 and Product Hunt Nov 10.
The part I care about most is checking whether the output is good: evaluation sets, an LLM judge calibrated against human labels, and a cost per session I can defend. The backend is public.
What it taught me. Generating a convincing itinerary is easy; building one I can trust is the work. Also that the backend is a good test bench for ideas from courses. After DeepLearning.AI’s Building Adaptive AI Agents I built the code graph it suggests (MiniLM for semantic starting points, PageRank over imports and co-edits) and ran a coding agent with and without it on two old bugs. On the simple bug the graph version used 5.5% more input tokens; on the harder one, about twice as many, and half a minute longer. The repo is small enough that the agent finds the files on its own. The idea may still pay off in a larger codebase, but the graph needs its own evaluation first.
Everyday skills active
Small automations for a family, built with Claude and connected tools. A grocery list that fills itself from what anyone says during the week; a wardrobe that picks the outfit.
The first one started with a family problem: remembering everything before the weekly grocery run. Now anyone can say “add milk” and it lands in a shared list, checked against the products we buy; on Friday a scheduled workflow fills the store cart. Getting the pieces to work together reliably took more effort than the idea, which is usually where the learning is. And yes, at some point I will probably put a camera in the fridge and let AI figure out what is missing before I do.
What it taught me. Access beats clever prompts: the automation only became useful once it could write to the list and the cart, not just talk about them. The prompts themselves I write on a walk, one or two miles talking to the model like a friend, then asking it to turn the conversation into a plan.
Sport Python learning active
Python notebooks for kids who love sports, on open NFL data, in Google Colab. Nothing to install.
The NFL Big Data Bowl on Kaggle is for grown-up analysts, so I made a simpler version for my ten-year-old, about the players he already knows. In the first notebook he checks whether the fastest players at the Combine become the best receivers. In the second he plays stat detective: best games, best seasons, a mystery player. He learns Python and how to look at data, and it is fun because it is about his favorite players. More NFL challenges next, then other sports. Ideas and new labs welcome.
What it taught me. Start from what the learner already loves and the syntax takes care of itself. He did not ask what a DataFrame is; he asked whether the fastest Combine runner became the best receiver, and found out.
ScoreZone done
A live sports scores app (NBA, MLB, golf) with standings, injuries and news. Built in a one-hour session with my son, from his idea.
He can imagine it. We can build it. The app was his idea and his requirements; the hour was the two of us at one keyboard.
What it taught me. A kid with a clear picture of what he wants is a better product owner than most specs. My job in that hour was to type less and ask more.