Medical bills are deliberately opaque. What if you could upload a bill or an EOB and get a plain-English explanation of what you’re actually being charged for - plus a draft appeal letter if something looks wrong?

The appeal letter generation. I built it as a secondary feature, almost an afterthought, but it turned out to be the thing people actually needed. Nobody wants to understand their bill - they want to fight it. The “decode” is really about giving people the language to push back.

Everyone says “tailor your resume to the job description,” but nobody actually does it because it takes forever. Could AI do the tedious part - matching your experience to their language - while keeping your voice?

The MatchMeter score. I added it as a quick pre-check - how well does your resume already fit this job? - and it became the most useful part. Some people realized they were applying to the wrong jobs entirely. The rewrite was less important than the honest assessment.

Policy organizations track hundreds of bills across federal and state legislatures. Most of that work is manual - staffers reading summaries, flagging relevance, writing briefs. Could AI do the triage so humans could focus on strategy?

The scoring model. I expected relevance matching to be straightforward, but the interesting problem turned out to be multi-dimensional impact scoring. A bill’s text might not mention your industry at all, but its downstream effects could reshape it. Teaching the model to see around corners was the real challenge.

Invasive species move through waterways, hitch rides on commercial vessels, and exploit gaps in infrastructure. The data exists - USGS sensors, vessel tracking, barrier locations - but nobody was fusing it into a single intelligence picture. Could AI connect those dots?

The Great Lakes Fishery Commission’s sea lamprey barrier data changed the whole project. When I overlaid barrier locations with vessel traffic patterns and species sighting data, risk corridors emerged that nobody had mapped before. The AI wasn’t discovering new data - it was seeing relationships humans hadn’t connected.

A Wall Street Journal investigation revealed how China is pushing into the Arctic alongside Russia - building military proximity to both Europe and America through a region most people still think of as empty ice. I read that and thought: the signals are all there - vessel traffic, ice conditions, port infrastructure, military posture - but nobody is fusing them into one picture. What would an OSINT platform look like that modeled Arctic operational viability in real time?

How much of operational viability comes down to timing. The individual signals - ice thickness, port accessibility, vessel positioning - are each well-understood. But the window where they all align is narrow and shifts unpredictably. The platform’s value isn’t in any single data stream; it’s in showing when the window opens and who’s positioned to exploit it.

“Zero to One” argues that your worldview - optimistic or pessimistic, determinate or indeterminate - shapes your outcomes more than talent or luck. Silicon Valley the HBO show made that thesis feel real in a way the book couldn’t. Could you take it further and make it playable? What if it was an arcade game?

Determinate pessimism is more fun to play than I expected. The fortress strategy - ration, defend, build walls - creates a compelling gameplay loop even though it leads to low outcomes. Players kept choosing it because it felt safe. That’s the whole point, made visceral.

March Madness bracket pools are one of the few things that make millions of people care about probability for three weeks a year. Could I build a real-time bracket app that handles the full tournament - 68 teams, play-in games, live scoring - as a full-stack exercise?

The cascade logic. When you change a pick in round 1, it has to propagate through every downstream round and invalidate affected picks. Getting that right - the “undo” problem - was harder than anything else in the app. It’s a tree traversal disguised as a sports app.

Could AI meet someone in a hard moment - not with advice, but with the right scripture and a beautiful image? Not a search engine for Bible verses, but something that listens to what you’re going through and responds with care.

The watercolor illustrations changed everything. I expected the verse matching to be the hard part, but it was the pairing of words with a generated image that made people pause. The combination felt less like a chatbot and more like receiving a handwritten note.

The PHQ-9 is one of the most widely used mental health screening tools in the world, but it’s a cold, clinical checklist. What if AI could walk someone through those same questions in a way that felt more like a conversation than a form?

How much the framing matters. The same nine questions, asked differently, surface different answers. People reflected more honestly when the questions didn’t feel like a test. That’s not a technology insight - it’s a human one.