Drone delivery uses Jev for twice-per-second flight decisions
The LLM plans the route once before takeoff, while Jev reads state twice a second and returns typed choices like CLIMB, STRAFE L, or HOLD, waking the LLM only on low confidence; the full run took 47 decisions with no text generation and cost about $0.0019.
a drone flew a full courier run and nobody asked an LLM what to do next 47 decisions. 0 words generated. $0.0019 for the whole run, in this demo the loop is split into four jobs: > LLM - writes the route once, before takeoff > JEV - reads the state 2 times a second and returns one typed choice: CLIMB, STRAFE L, HOLD > GATE - confidence of 0.85 or higher goes straight through, anything lower wakes the LLM > CODE - actually flies the thing eight fields come in: altitude, wire clearance, wind, battery, drop zone distance, dog distance, window, payload. one label comes out in about 200 ms wires are close and altitude is low, so it climbs. payload is empty and the zone is reached, so it holds. none of that needs a paragraph that's why the cost stays near zero. the expensive model only shows up when the state is unclear planning is slow and expensive. steering has to be fast and cheap. splitting them is the whole trick state schema, gate logic and the full loop in the thread below