Aidan Soares

City Insight started with a question I kept asking: is there a tool that compares cities using both hard data and what residents actually think about living there? I could not find one that did both well, so I built it.

The interesting problem was not the data: it was the query interface. Most location-comparison tools force you to choose metrics from a predefined list, which means they answer the question the designer anticipated, not the one you actually have. I built an AI chat interface on top of the structured dataset so that natural-language questions resolve against the same data driving the radar charts and maps. The challenge was keeping every interface synchronized: a filter set via chat should immediately be reflected in every view at once.

Wildfire Command is a browser-based simulation where players manage real-time wildfire response, deploying helicopters, trucks, and crews against procedurally generated terrain. I had worked in wildland fuel reduction before this, so I had a specific sense of what plausible fire spread should actually look like.

The core problem was making that behavior feel real without the simulation being prohibitively expensive. Each cell in the terrain grid maintains wind exposure, fuel load, and moisture state; spread probability is recomputed only on changed cells, not the full grid. The result is fire that reads as coherent: it runs uphill, stalls on rocky outcroppings, backs slowly into the wind, without blocking the main thread. Terrain itself is procedurally generated using Perlin noise for elevation, BSP partitioning for large-scale structure, and cellular automata to produce natural-looking fuel density variation.

Exercise Tracker is a workout logger: create, view, edit, and delete exercises with name, reps, weight, unit, and date. The goal was to build something complete and correct rather than clever, and to understand the full MERN stack end to end by building something I would actually use.

The decision I spent the most time on was weight units. Rather than storing a normalized value and converting on display, I store the weight and unit as entered. This means the data model reflects what the user typed, there is no lossy conversion happening in the background, and the validation and edit logic has to be unit-aware throughout. That last constraint turned out to be the most interesting part of the project.