Work
Grouped by the kind of problem, not by client. Each write-up covers the actual problem, the approach, the trade-offs, and what it changed — including where the evidence stops.
A nonprofit run club had to wait on an agency every time its site broke. Rebuilt on hosting the organization controls, with measurement that works from launch and a handoff a volunteer can actually run. Mobile performance went from 65 to 94.
Sole designer and developer. The organization is not named here, at its discretion.
A tool that answers "which county is this ZIP code in?" correctly — including the messy cases where one ZIP crosses several county lines. Built so an AI assistant can look the answer up directly instead of someone checking by hand.
Personal project, built and shipped by Ismael Melendez.
Five candidate datasets audited and rejected before any analysis was written, and a sample-size bar set in advance so an early-looking pattern couldn't argue its way in later.
Personal project, in progress. Sole designer and developer.
A Panda3D maze shooter written in college in 2016, rebuilt for the web twice — once in 2D Canvas, once in 3D with the original art and its 48-joint character rig. Both play on the original's own maze, extracted from the 3D model rather than redrawn.
The interesting part isn't the graphics. A port is a claim that the rules still behave the same, and that claim is easy to get wrong and hard to notice. So the rules are tested: 29 checks run the real game module headlessly — a donut is always eaten, the portal refuses you without the quota, orbs alone are not enough — rather than someone playing for a few minutes and deciding it felt right.
Personal project. Sole developer. Play the 3D version, watch the tests run, or read the 2016 original. The method behind it is written up in how do you know a rewrite still works?
From prior nonprofit work
First-person write-ups of work done in previous nonprofit roles, before Soft Science Studio. They describe what I did, what the evidence showed, and where it stopped short of proving something.
Search Console research across five search topics, collaborating on page addresses and headings, and restoring useful pages lost in an earlier redesign.
A 2024 influencer campaign analysis combining social-platform and website data — including the finding that traffic and engagement moved in opposite directions.
Requirements gathering, feature comparison, staff working sessions, and following up on whether data actually moved between systems.
Redesigning intake questions so fundraising colleagues got the context they needed alongside attendance numbers.
If something in your data doesn't line up as cleanly as everyone assumes, that's usually where the real work is.