I know $90 for a dining app sounds absurd. And it’s not even an app; it's a wrapper for an LLM. But it is how it is built that makes it so expensive. I had to make a whole API just to call it (which because its own SAAS). All just to bypass dumb google reviews.

Google Maps and reviews send people to places optimized for tourists and good copywriters. To find actual local hole-in-the-walls algorithmically, I had to first build that api (called BWENDI), a "spatial gravity" engine using 100GB+ of tweaked OSM, GeoNames, and other proprietary data. Instead of aggregating reviews, it mathematically calculates foot-traffic, throughput, transaction stats, and economic criticality among other factors.

Bwendi is A Python/Node ETL pipeline feeding an LMDB-backed context API. It uses a proprietary 1MB binary grid served via Cloudflare Workers for millisecond edge reads with near-zero overhead, hosted in Switzerland.

This was done of course to get the purest location context around every street or neighborhood, so that the LLM stack could breathe and think properly.

So far the results are good.

Rip apart my edge architecture, the gravity math, or the aggressive pricing. I’m here for the roast.