Manav Khosla
Computer Science and Business Adminstration Student at the University of Michigan, College of Engineering and Stephen M. Ross School of Business.Projects listed below... feel free to reach out at [email protected]!
Gunshot Detection Alarm
A 3D-printed, Raspberry Pi 4-powered IoT device that uses machine learning to detect gunshots in real time, instantly alerting local authorities.
SafeStreets
A cross-platform crime-mapping app that enables real-time incident reporting between users, keeping communities aware and safe.
Scribe
A mobile OCR app that uses deep learning and computer vision to convert handwritten documents into accurate, legible digital text for students and educators.
Gunshot Detection Alarm
A 3D-printed, Raspberry Pi 4-powered IoT module, this Gunshot Detection Alarm is engineered for real-time audio surveillance, integrating a low-cost MEMS microphone, onboard preprocessing, and a Convolutional Recurrent Neural Network (CRNN) pipeline. Audio streams are captured in rolling windows, transformed into mel spectrograms, and passed through CNN layers for spectral feature extraction followed by RNN units for temporal context modeling. Furthermore, this device leverages Inverse Frequency Weighting to mitigate class imbalance in sparse gunshot datasets and applies data augmentation strategies such as random cropping and time-shifting.The case includes a combination of TPU + PLA filament, and contains isolated chambers for the Raspberry Pi 4, the USB microphone, and the USB speaker. Finally, detected gunshots are paired with mobile endpoints via a React Native interface, with local authorities and individuals notified through Twilio SMS messages. With a $300 build budget, this project achieved up to 95% classification accuracy.Full source code and implementation details are available on Github: https://github.com/khoslamanav25/GunshotDetectionAlarmA live demonstration of the system in operation can be viewed here: https://youtube.com/shorts/2qr7kStExkM?si=VYyx8P9yHGdpnPtJ
SafeStreets
SafeStreets is a cross-platform mobile application built with React Native, designed to provide real-time crime awareness through a dynamic heat map, live reporting, and data integration. The system leverages Selenium to web scrape NYPD CompStat datasets to process official crime statistics, visualized via the Google Maps API as an interactive heat map with time-based filtering (24 hours, 1 week, 1 month, etc.). The app also integrates the NYTimes API, allowing a dedicated news feed to showcase dangerous events in the area around users. Google Firebase also manages user authentication and the Google Firestore database process image, caption, and all other user data.Furthermore, SafeStreets incorporates a live crime reporting pipeline where users can upload images of incidents. These are processed through a CLIPxGPT image-captioning model tuned on the FLickr30k dataset, generating captions that contextualize visual evidence. Each report is geotagged and displayed on the heat map, and when a high-threat event is dedicated, nearby users and local law enforcement are alerted, with corresponding location metadata attached.Full source code and implementation details are available on Github: https://github.com/khoslamanav25/SafeStreetsBronxNYA live demonstration of the system in operation can be viewed here: https://youtu.be/zTUIvT2SCo8?si=k4htoRjbJAnEjWlq
Scribe
Scribe is a full-stack OCR React Native mobile application, designed to digitize noisy handwritten inputs through a multi-stage preprocessing and deep learning framework. First, OpenCV and scikit-image is utilized to apply eight sequential image normalization stages: grayscale conversion, adaptive thresholding, skew correction, dilation/erosion, segmentation, and contour analysis. Once preprocessed, images are passed into a CNN implemented in TensorFlow and trained on the EMNIST-Balanced dataset, which contains 131,000+ labeled samples across 47 classes. Stratified sampling techniques and input upscaling were used to mitigate class imbalance, yielding transcription accuracies exceeding 91.5% on unseen handwritten text.This model is fully integrated into a React Native application. Authentication and data records are handled via Google Firebase Authentication and Firestore DB/Storage, enabling users to log in, upload handwritten notes, and retrieve transcribed outputs across devices.A live demonstration of the system in operation can be viewed here: https://youtu.be/AYRvUX19s1k?si=r-zM6hD3Xsw-fU9l