I build the systems underneath, and solve the hard problems.

Ask a language model whether a trail is melted out this weekend and it answers from the season rather than the year. This one checks. It pulls the hard and soft constraints out of a natural-language query, then runs a cost-ordered tool chain over SNOTEL snow depth, NWS forecasts, USFS road status and NPS closures — and returns ranked trails together with the values it actually looked at.

A local's guide to Seattle, built on top of my own Google Maps history. Google has no API for saved lists, so an importer reads a Takeout export, lifts coordinates straight out of the saved-place URLs, geocodes the stragglers, and filters out everything personal before any of it ships. Every list opens onto a map, with search across all of them at once.

Enter a birthday and scroll a century of world events, each one stamped with how old you were when it happened. The events are not the point — you know most of them. What lands is the number beside them, which makes the date arithmetic the whole product: off by one anywhere and the effect collapses. Zero dependencies, no build step, 49 tests behind the age maths.

An autonomous-vehicle simulator built around pluggable driving rules: a common rule interface over a shared road, lane and vehicle model, with concrete strategies for holding a lane and for passing on the left or on either side. A Swing front end runs the scenarios live, so a rule is something you watch play out rather than read the output of.

Recovers a decryption key by evolving toward it rather than searching for it. Chromosomes encode candidate keys, fitness scores each one on how much like real language the output reads, and successive generations converge on plaintext. Written for Northeastern's algorithms course, with the analysis and results written up alongside the code.