Computer Vision & IoT

A hospitality group operating 12 properties across South India was losing an estimated 18% of their food budget to waste — roughly INR 2.4 crore annually. They had tried manual logging (asking kitchen staff to weigh and record waste in spreadsheets), but compliance was poor and the data was unreliable. They needed an automated system that could identify food types, estimate quantities, log waste events to a cloud database, and generate actionable reports for management — all without disrupting the fast-paced kitchen workflow.

The secondary project was a face recognition system for employee attendance and restricted-area access control. The existing biometric system (fingerprint scanners) was failing frequently due to wet or greasy hands in the kitchen environment, and buddy-punching (one employee clocking in for another) was costing the company thousands of hours of ghost attendance annually. They wanted a contactless, camera-based system that could identify employees in under two seconds, even in variable lighting conditions, and integrate with their existing HR software.

Both systems needed to run on edge devices (no cloud inference for latency and privacy reasons), work reliably in challenging environments (steam, variable lighting, partial occlusions), and provide real-time alerts through Telegram and Slack when anomalies were detected — such as unusually high waste volumes or unauthorized access attempts.

For the food waste tracker, we trained a custom YOLOv8 model on a dataset of 15,000 labeled images covering 47 food categories common in South Indian hospitality (rice varieties, curries, breads, fruits, prepared dishes). The model ran on NVIDIA Jetson Orin Nano devices mounted above waste bins, processing frames at 24 FPS. When food was deposited, the system captured a burst of frames, ran inference to identify food types, estimated volume using depth data from a stereo camera pair, and logged the event with a timestamp, food category, estimated weight, and a reference image to a Firebase Realtime Database.

The face recognition system used a two-stage pipeline: MTCNN for face detection and alignment, followed by ArcFace embeddings for recognition. We enrolled 340 employees across all properties, storing their face embeddings in a local SQLite database on each edge device for sub-second lookups. The system handled variable lighting by applying histogram equalization and contrast-limited adaptive histogram equalization (CLAHE) as preprocessing steps. For anti-spoofing, we implemented a liveness detection module that analyzed texture patterns to distinguish real faces from printed photos or screen displays.

Both systems reported to a centralized dashboard built with React and Chart.js, showing real-time waste trends, attendance records, and alert histories. Notifications were dispatched through a Node.js service that integrated with Telegram Bot API and Slack Webhooks. The entire edge-to-cloud pipeline was designed to work with intermittent internet connectivity — events were queued locally and synced when the connection was restored, ensuring zero data loss even during network outages.