Projects

Worth noting: The OCR targets the collector number rather than the card name — numbers are small, isolated, and uniform, while card names have art interference and stylized fonts. That one decision made the pipeline dramatically more reliable, and means the app works on any card regardless of language.

Worth noting: Snowflake queries take 1–2 seconds cold. Running them live on every request would make the site unusable. The solution was a TTL cache with a background pre-warm thread that fires 3 seconds after deploy and hits every heavy endpoint. First real user always gets a cached response. It's a small thing, but it's the difference between a demo that works in a screencast and a product someone actually uses.

Worth noting: All the signal processing happens in the browser — no backend audio pipeline, just the Web Audio API's AnalyserNode feeding FFT data into shader uniforms 60 times a second. The GPU handles 60,000 particles per frame; the CPU handles the band math and beat detection. The constraint that made the whole thing interesting: tab audio capture via getDisplayMedia lets the visualizer react to whatever's already playing in another tab, with no system audio driver required.

Worth noting: SVI and PLACES are both well-maintained public datasets, but they weren't designed to be joined. Getting the geographic grain right across 73,000 census tracts took longer than the analysis itself. Once that was clean, the 1.64x diabetes finding came out clearly — which is either obvious or alarming depending on who you ask.

Worth noting: The interesting part isn't the generation — it's that once the prompts are parameterized and the queue is set up, there's nothing left to babysit. Runs on local GPU, files organize themselves, no per-image API costs accumulating in the background.

Worth noting: The 200x stat is real. What made it clean was scoping narrowly — this was a migration, not a redesign. The main risks were DAX parity with existing MDX logic and partition strategy. Those are where most of these go wrong.

Worth noting: At this scale, the query logic isn't the hard part. What makes a pipeline actually reliable is how it fails — whether it surfaces errors loudly and specifically when something upstream changes. Silent corruption in financial data is worse than downtime.

Worth noting: The requirements conversation is most of the job. People know the data they want to see; they're less sure what question they're actually trying to answer. Getting that right upfront is the difference between a dashboard that gets used and one that gets opened once.

Worth noting: Performance problems in Tableau usually concentrate in a few specific places — extract size, custom SQL that reruns on every page load, LODs that don't need to be LODs. Telemetry makes it obvious which. The 20% improvement came from fixing those, not a blanket cleanup.

Worth noting: The main issue wasn't the software — it was that updates required remembering to ship them, so they rarely happened. Moving to web made deployment automatic and removed the per-machine installation problem. The 15% productivity gain is mostly that friction gone.