Headroom

Headroom: more EV chargers on the grid connection a site already has Offices, malls and depots share one meter between the building and its EV chargers, and US utilities bill the highest 15 minutes of each month. Cars arrive when the building is busiest, so one 50 kW spike at 6 pm can add $2,572 to a single month's bill. Owners stop adding chargers, or wait years for a bigger grid connection. Headroom is an AI operator that fits more chargers on the connection a site already has. Every 15 minutes it: 1.forecasts the building's load, with a calibrated worst case 2.plans the charging in the room the building leaves, using a linear program 3.passes the plan through a deterministic safety layer that nothing can override 4.sends it to the chargers as OCPP 1.6J power limits Meta's Muse Spark runs the site day to day. It reads the utility's tariff, which is accepted only if it reproduces the real bills. It takes requests by text or by voice through Muse Voice Transcribe, spots blocked bays on camera, and answers questions like "can I add 12 more chargers?" from simulation. It can act only through validated tools, and every number in its answers is traced to a tool result. If the internet drops, Muse Glimmer takes over on the site's own GPU. Charging never depends on any model. Results on real data (Palo Alto charging sessions, a real metered building, the utility's own tariff, on months the model never saw): · 48 chargers instead of 36 on the same 1,000 kW connection, with 100% of drivers' energy delivered and 0 times over the limit. · 32% lower demand charges than standard smart charging ($2,889 vs $4,231 a month), which is 82% of the saving a perfect-foresight controller gets. · Across 49 real buildings at 7 US sites: more chargers in 77% of them, fewer in none. · On a second utility's tariff (PG&E B-19): 18% lower demand charges. · Over the OCPP 1.6J charger protocol, 480 of 480 power limits were obeyed. Meta models used: 1.Muse Spark 1.3: the operator. 2.Muse Voice Transcribe 1.0: spoken requests. 3.Muse Glimmer: the on-site backup. Every call is recorded and replayed in the live demo, and total model spend was $0.60. The project is open source under the MIT licence. Links: · demo video - https://youtu.be/1-Ch1mL8464 · live app - https://headroom-ev.streamlit.app/ · walkthrough - https://headroom-meta.vercel.app/demo/ · code - https://github.com/SahanaGaneshvel/META