Nightline

Utilities see the grid go dark. Nightlight sees who's still in the dark: a $15 ESP32 box reports indoor heat and SOS after a hurricane, and NVIDIA Nemotron helps crews reach the riskiest homes first.

Comment

Inspiration

When Hurricane Ian hit Southwest Florida in September 2022, about 2.6 million Florida customers lost power, and NASA's night-light satellites showed whole neighborhoods around Fort Myers go dark. Utilities could see which feeders were down. They could not see which homes had an 80-year-old on an oxygen concentrator, sitting in rising heat with no way to call for help. Restoration usually follows the grid: biggest feeders first. We wanted to restore by who is in danger.

What it does

Nightlight is a small box that keeps reporting after the power goes out:

  • Wall power on or off
  • Indoor heat index (temperature and humidity)
  • An SOS button
  • A medical-load flag (a flag only, never a diagnosis)

Every 5 seconds it sends a 24-byte packet over a phone hotspot. If the network drops, packets wait in a queue on the device and send in order when it comes back, so the outage timeline has no gaps.

In the cloud, a deterministic scorer ranks every home by danger: heat index and its trend, hours dark, time since anyone moved, the medical flag, occupants, and SOS. It also forecasts the hours until a home becomes unsafe. NVIDIA Nemotron then explains the ranking and drafts the crew's restoration brief. It never decides the order itself, and a dispatcher approves.

The console has three views: a map of ranked homes, a device view with live charts and the raw packets from the real box, and a crew view with the Nemotron brief and resident check-ins.

How we built it

  • Hardware: ESP32 DevKit V1 + DHT11 temperature/humidity sensor + a pushbutton, wired on a breadboard with no soldering. A jumper wire stands in for the wall-power sensor in the demo. The firmware (Arduino C++) packs a frozen 24-byte packet with a CRC, runs a 900× demo clock, and keeps an offline queue.
  • Cloud: Python 3.11 + FastAPI + SQLite. The packet format is defined once and matched byte for byte across docs/PACKET.md, the firmware header and the Python decoder, with a round-trip test.
  • AI: NVIDIA Nemotron 3 through Nebius Token Factory (OpenAI-compatible API).
    • Nemotron 3 Nano (30B-A3B): one-sentence explanations for the top 20 homes, run in parallel, plus parsing messy resident messages ("my mom's oxygen machine battery dies tonight") into structured needs.
    • Nemotron 3 Super (120B-A12B): the crew restoration brief, built from the scorer's ranking.
  • Real data: HHS emPOWER (electricity-dependent Medicare beneficiaries per ZIP) and U.S. Census ACS 2023 (households with someone 65+), for 937 Florida ZIPs. Our demo area, Pine Island (ZIP 33956), has 1,852 households, 55% with someone 65+, and 51 people on powered medical equipment.
  • Console: plain HTML/JS with Leaflet, GSAP motion and Lucide icons, in light and dark modes, with no build step.

What's real and what's simulated: one real unit (our ESP32). The other 150 homes on the map are simulated from the real Pine Island ratios and labeled SIMULATED everywhere.

Challenges we ran into

  • First time building hardware. Our ESP32 has a single 3.3 V pin, so we shared it through a breadboard column. We also had to work out which DHT11 pin is which, and which legs of a 12 mm button are connected.
  • Nemotron 3 Nano returned empty answers. It reasons by default, and with a small max_tokens it spent the whole budget thinking. Turning thinking off for the Nano calls fixed it: answers came back in about a second.
  • Honesty about results. Our first outage simulation showed only about a 2% improvement for danger-first restoration, because it spread vulnerable homes evenly across feeders. Real data says otherwise: the share of residents on powered medical equipment ranges from 0% to 64% across Florida ZIPs. We report the small number honestly and name the assumption, instead of tuning it.

Accomplishments that we're proud of

  • A real device sending real packets end to end, decoded by the cloud byte for byte.
  • A ranking that is deterministic and explainable, with the LLM kept in the explainer role.
  • Every simulated element clearly labeled, and every statistic sourced.

What we learned

How to wire and flash a microcontroller from zero, how to design a tiny binary protocol that survives bad networks, and how to use an LLM where it helps (explaining and drafting) without letting it make the life-or-death call.

What's next for Nightlight

A real wall-power sensor instead of the jumper, a motion sensor, battery reporting, a pilot with a county emergency office, and deploying the API as a Nebius Serverless Endpoint for always-on ingest.

Built With

Share this project:

Updates

Submission history