Vaibhav
Goel.

A dual-stream 1D-CNN with a physics-based residual filter, tested against eleven other models under leakage-free protocols on the CWRU and Paderborn datasets. The ~99% accuracies common in the literature collapse once the test bearings are truly unseen.

• 4 protocols, 2 datasets, 12 models
• 0.30 ms INT8 inference on CPU

Admins upload call recordings; Whisper transcribes them and Llama 3.3 scores quality, CSAT and sentiment. Agents self-evaluate and HODs rate compliance. Audio only streams through JWT-authenticated blobs and is purged on a schedule.

• Production move from SQLite to PostgreSQL
• 3 roles, no public audio URLs

Transcribes banking calls with Whisper, runs a deterministic keyword rule engine for intent, priority and risk, masks account and card numbers, and has an AI verifier review each decision, with the rules standing whenever the AI fails.

• Rule-based fallback for every decision
• Masked account numbers, API-key auth, rate limits

Fuses five sensors on an ESP32-S3 and labels each reading normal, warning or fire with safety rules and a Random Forest trained on 367,550 real sensor rows, streamed to a live Dash digital twin.

• 94.93% test accuracy
• 98% fire recall

Labels every pixel of a road scene as road, vehicle, pedestrian, sidewalk, traffic sign or background. Trained at full 1024×2048 resolution on Cityscapes and the India Driving Dataset together, a pretrained SegFormer-B2 reaches 0.867 mIoU on held-out German streets and 0.868 on held-out Indian roads, up from 0.679 before IDD training. Grad-CAM explanations and validated ONNX export.

• 0.867 mIoU on Cityscapes
• 0.868 mIoU on Indian roads