Splitwatts

SplitWatt turns roommate electricity arguments into a fair split by letting everyone describe what they actually used. It also sniffs out who burned through the prepaid units too fast so the real energy vampire gets called out.

Live demoGitHub0

Overview

Shared electricity bills create constant friction in student housing because no one has clear proof of actual usage. SplitWatt lets roommates describe their habits in everyday language, then uses AI to estimate consumption and produce a transparent, proportional split of the total bill. A second mode helps anyone who suspects prepaid units disappeared too quickly by comparing their own estimated use against the amount loaded and returning a clear verdict.

How it uses Vultr

How Vultr was used in SplitWatt SplitWatt runs entirely on Vultr. The backend lives on a Vultr Cloud Compute instance (Ubuntu 22.04, plan vc2-1c-1gb in Seattle). Nginx on that instance reverse-proxies port 80 to a Node.js + Express server that serves the single-page frontend and handles the two API routes. The AI layer uses Vultr Serverless Inference with a LLaMA model. Every time a user describes their electricity usage, the Node.js server sends the description to Vultr’s OpenAI-compatible endpoint (https://api.vultrinference.com/v1). The model returns an estimated kWh figure (and a short breakdown) for Split Mode, or a full investigation verdict for Investigate Mode. The inference API key is stored only in a .env file on the server and is never exposed to the browser. Why Vultr We chose Vultr because it let us go from zero to a live, AI-powered product in one place. Cloud Compute gave us a cheap, ready Ubuntu server we could SSH into and deploy with nginx and pm2 in minutes. Serverless Inference gave us a working LLM without provisioning GPUs, managing model weights, or writing custom inference code — we just pointed our existing Node.js fetch calls at their OpenAI-compatible API. The $200 HackGlobal credits covered both the Compute instance and all inference usage for the entire weekend, so we could focus on the product instead of infrastructure costs or credit limits. The combination also kept the architecture simple: one server for the app, one managed inference endpoint for the AI, and a clean separation so the frontend never sees secrets.

No comments yet

Be the first to comment on this project.