Trash Panda & CurbFusion

Houston publishes hundreds of datasets that touch a curb — parcels, 311 calls, garbage pickup days, traffic counts, flood
zones, bus stops — and none of them answers a question on its own. Planners, delivery drivers, homebuyers, and city crews
each need a different answer, and each has to cross-reference tables by hand.

CurbFusion pulls eighteen public layers from City of Houston, HCAD, METRO, TxDOT, and FEMA endpoints (no keys), snaps them
onto one grid — block face × hour of week — and exposes every dataset as a 0–1 component, down to individual 311 case
types. A recipe is a set of signed weights plus a time window; scoring is a dot product in the browser. Ten recipes ship,
a Recipe Lab lets anyone drag their own, and a Recipe Agent turns a plain sentence into a proposed recipe with a reason per
dataset and a wishlist of data it couldn't get. We proved the framework is need-agnostic with the hardest customer we could
think of: a raccoon family, scored on garbage nights, tree cover, dead-animal calls, and quiet streets.

Impact: one day of data compilation, then a new question costs a sentence. Nine Houston corridors today; the pipeline takes
any bounding box. The datasets the agent asks for — meter transactions, citations, rideshare pickups — are ones the city and
its partners already hold, which turns the model into measurement.