Working tools, not slide decks.

Most organizations use email to communicate with customers, and mistakes are expensive. A marketing email that lands in spam, renders broken on mobile, or triggers compliance flags doesn't just fail to convert. It damages trust.

The problem is that most teams don't find out until after they've sent to their entire list. Testing tools exist, but they're typically buried inside enterprise platforms or require technical setup that marketing teams don't have time for.

EasyEmailTest is a standalone QA platform built for speed. Upload an email (raw file, HTML, or forwarded message) and get back an instant analysis: spam risk score, rendering previews across devices, content recommendations, and compliance flags. No accounts to configure, no integrations to maintain.

The platform handles the full stack: user authentication, email processing, AI-powered content analysis, and structured reporting. It's live in beta (working through scenario tests).

What it demonstrates: End-to-end product development: from identifying a real workflow gap to shipping a working system with billing, authentication, and email infrastructure.

Predictive models are usually locked away: buried in academic papers, hidden inside proprietary systems, or stuck in spreadsheets that only their creators can use. The gap between "interesting analysis" and "something people can actually interact with" is where most data projects die.

Fútbol Time Machine takes two models, a classic Elo rating system and an enhanced machine-learning model trained on 19 features (including FIFA rank, squad value, and recent form), and makes them explorable. Pick any two national teams, any year, and see the simulated outcome: win probability, expected goals, the historical context that shapes the result.

Neither model is novel. Elo systems are well-established in sports analytics. What's different is the delivery: a public-facing web application where anyone can play with scenarios, test their intuitions, and understand how the results are produced. No login, no paywall, no data science degree required.

Building it required the full ML deployment pipeline: data collection and cleaning, model training and validation, API design, frontend development, and infrastructure that stays up without babysitting.

What it demonstrates: Machine learning isn't useful until it's deployed. This project is proof of the complete journey: from raw historical data through trained model to interactive public application.

My response was to minimize the overhead required to balance working, building, and staying up to date on technologies and their day-to-day applications. The "context switching" tax had me going in too many directions and keeping too much in my head.

I built internal tooling: systems that handle the repetitive knowledge work so I can focus on the decisions that actually require judgment.

One tool maintains a knowledge wiki: drop in a document or URL, and it synthesizes the content into my existing notes, handling the formatting, cross-referencing, and organization automatically.

These aren't chatbot wrappers. They're structured systems with guardrails that keep the goal in view and the work on course. Claude has done most of the development work. Goose, an open-source, general-purpose agent, handles all of the development operations. My worst grade ever was a D in computer science. Twenty years of project work taught me the part that matters here: set the goal, then hold the work to it.

What it demonstrates: I set the goal and the definition of done, and AI agents do the building while staying on course. That's project management applied to software.