This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle

What I Built

Outside college campus gates in India (like ITER/SOA, KIIT, and OUTR in Bhubaneswar), late-night roll counters, burger stalls, and small cloud kitchens rely heavily on hungry hostel students. However, their razor-thin margins face a daily threat: unpredictable perishable inventory waste.

Freshly marinated chicken, paneer cubes, and burger buns spoil overnight under humid coastal conditions. If it suddenly rains at 9:00 PM, or if hostels enforce quiet-hour study gates before mid-semester exams, stall owners routinely throw out ₹1,000 to ₹2,500 worth of fresh stock by morning.

HostelBites is an autonomous inventory liquidation copilot built for the daily 8:00 PM kitchen check-in:

1. The 8:00 PM Inventory Audit: The line cook or stall owner enters remaining raw stock weights alongside environmental conditions (exam week flags, rain, day of the week).
2. Zero-Shot Tabular Spoilage Prediction (TabPFN): Ingests the tabular kitchen metrics to classify spoilage risk, outputs posterior class-probability distributions, and calculates continuous clearance windows and mandi financial loss exposure.
3. Campus-Targeted Flash Combos (Gemma 2): Uses gemma2-9b-it via Groq to synthesize high-margin flash combo meals that bundle surplus stock, generating ready-to-broadcast WhatsApp copy tailored with local campus slang and exam urgency.
4. Hands-Free Kitchen Audio Briefing (ElevenLabs): Converts the preparation priorities into a clear Hinglish voice walkthrough using the Turbo engine, letting the line cook hear batch-fry priorities over loud kitchen exhaust fans.

Stall owners gladly pay a modest ₹199/month micro-subscription because salvaging just one evening's worth of unsold chicken pays for the service for an entire semester.

Demo

• Live Application: https://hostelbitesapp.vercel.app

Code

The entire open-source repository—including the tabular prior dataset, TabPFN foundation pipelines, Gemma 2 prompt engineering, and the Tech Editorial frontend—is public on GitHub:

How I Built It

HostelBites links tabular machine learning with generative copy and neural speech synthesis:

• Prior Labs TabPFN-3.5 Foundation Model: Rather than training a brittle, heavy gradient-boosted tree that requires thousands of samples, TabPFN functions as an in-context tabular foundation model. Fitted directly over 150 nights of campus kitchen records (data/midnight_orders.csv), it delivers:
  • Categorical Risk Classification: Classifies risk into Safe, Moderate, High, and Critical.
  • Calibrated Posterior Probabilities: Provides confidence percentages across every risk tier.
  • Continuous Clearance Windows: Projects remaining liquidation hours before demand drops to zero.
  • Outlier Detection: Flags dangerous inventory build-ups (e.g., surplus poultry during sudden storms).
• Gemma 2 (gemma2-9b-it via Groq): Analyzes the inventory surplus and risk status to generate campus-themed WhatsApp broadcasts with emojis, gate delivery instructions, and combo names ("End-Sem All-Nighter Double Roll").
• ElevenLabs Neural Turbo (eleven_turbo_v2_5): Synthesizes practical 30-second Hinglish voice instructions with the George voice profile, backed by automatic Web Speech API browser fallbacks.
• Frontend Architecture & Hosting: Styled with a "Tech Editorial" aesthetic (Playfair Display, Space Grotesk, and Space Mono) on a 4-column guide grid with scan-line progress indicators, deployed on Vercel.

Why Does Open Innovation Matter?

Open-weight models and open foundation models are essential for micro-enterprises:

1. Protecting Micro-Margins: Proprietary closed APIs charge high token fees that consume the small margins of street vendors. Open-weight models like Gemma 2 and CPU-efficient foundation models like TabPFN run at near-zero inference cost.
2. Data Sovereignty & Local Control: Daily inventory ledgers, raw purchase costs, and local customer logs stay private within the vendor's application rather than being harvested by commercial cloud providers.
3. Zero-Shot Adaptability: TabPFN's transformer architecture evaluates small tabular datasets directly in memory without hyperparameter tuning, bringing institutional-grade yield optimization to grassroots developers.

My First User

To test HostelBites in the real world, I visited a popular roll and paratha stall near the engineering student hostels in Jagamara, Bhubaneswar around 8:30 PM.

The cook had approximately 3.5 kg of pre-marinated chicken and 30 dough balls remaining on a quiet evening:

• TabPFN Output: Classified the situation as Moderate risk, computing a 2.5-hour clearance window and ₹1,415 in at-risk stock.
• Gemma 2 Output: Formatted an urgent "Exam Midnight Craving Alert: Jumbo Butter Chicken Roll @ ₹119" WhatsApp broadcast.
• Feedback: The owner explained that refrigerating opened, marinated meat degrades its texture for the next day. He forwarded the broadcast into two local student hostel groups. Within 40 minutes, the stall received 9 orders, clearing over 2 kg of chicken before midnight. He affirmed that a ₹199/month fee is an easy investment if it reliably clears perishable stock during exam seasons or rainy days.

Prize Categories

• Best Use of TabPFN
• Best Use of Gemma
• Best Use of ElevenLabs
• Best Use of Vercel