Inspiration
We three work as Student Food Service Workers at the campus Dine-in center at IIT. Every shift we watch pounds of food get thrown away with no log, no trend, no signal. Just a full bin and a bag change. We built the tool that should have already existed.
What it does
HawkWaste is a two-screen web app — one for kitchen staff, one for management.
FSW Logger: Photograph the waste bin at bag change, select shift and date, submit. GPT-4o-mini vision estimates bin fullness and weight in pounds. No scales, no hardware, one tap.
Manager Dashboard: A week of waste data by shift, highest/lowest waste items, an AI recommendation card, and a menu cross-reference panel showing every item on today's menu with its predicted waste tier, lbs cooked, and risk level — surfacing high-risk items before service starts.
How we built it
Backend: Python Flask with five endpoints — photo analysis, logs, AI recommendation, menu cross-reference, and waste prediction — run locally
Frontend: React + recharts, custom design system with IIT Scarlet branding, running locally
AI: OpenAI gpt-4o-mini for image analysis and text (weekly recommendations)
Data: 464 real IIT Commons menu items extracted from dineoncampus.com across 6 days, plus a scored dataset where each item is rated on familiarity, cultural fit, visual appeal, comfort level, and protein presence
Total API cost for the full demo: under $5.
Challenges we ran into
Mixed bin problem: We photograph the post-service scrape bin, so there's no per-item breakdown from one photo. We solved this by using photo logs as the audit trail and the menu scoring system as the item-level signal — two separate layers that together tell the full story. AI accuracy: A bin photo is ambiguous. We reframed it: trend accuracy beats point accuracy. A reading 5% off still correctly shows whether waste is going up or down week over week. Real data from a binary format: Our weekly menu was in a .docx file a ZIP of binary protobuf .iwa files. We extracted, converted to .docx, and parsed 464 real menu items by hand.
Accomplishments that we're proud of
Zero hardware required just one phone camera, $5/month in API costs 464 real IIT Commons menu items structured with station, shift, and waste scores A menu cross-reference panel that tells management before service which items to cook less of A UI designed for two very different users: FSWs who need one tap in a kitchen, and managers who need dense weekly data
What we learned
FSW knowledge is a competitive moat. Knowing that IIT students don't recognize chilaquiles, that halal students skip the sausage station, and that cauliflower in any form consistently wastes 50%+ that insider context made every layer of the system sharper than any external dataset could.
What's next for HawkWaste
Phase 2: Fixed overhead camera at the scrape station, continuous automated logging, no human step Phase 3: Camera at tray return before scraping, item-level recognition matched against the daily menu for true per-item waste tracking Expansion: Every Chartwells campus uses dineoncampus.com and the same station structure, one new menu dataset per location to deploy anywhere
Log in or sign up for Devpost to join the conversation.