Mohit Angrish

Instance Segmentation On Aerial Images

• Worked on various use-case for getting outline for various structures(buildings, pool, trampoline) from the aerial images.
• Improved the recall for all classes, building from 41 to 98%, deck from 8% to 42%, pool from 80% to 96% and trampoline from 77% to 93%.
• Also identifying whether there are obstructions present on the building which may impact the outline accuracy.
• Improving the model detection of obstruction from 40% to 95%.
• Also improved the pipeline yield from 60% to 80% by changing the solution approach for the pipeline.
• Responsible for designing the solution for the use-cases, data acquisition, model training, model quantization for faster inference and make it deployable on triton server. Also performing failure case analysis to enchase the current pipeline for further improvement.

Computer Vision : Sewer Defect Detection : Video Precessing and Object Detection

• Lead team of 6 from to develop, train/test multiple Object detection models | classification models | OCR models identifying 74 visually separable defects occurring in sewage pipes and Distance counter of pipe respectively according to NASCO standards. Devised different strategies to verify existing marked defects and non detected defects.
• Closely Worked with UI and Application team to optimize workloads and add more product experience features to develop feedback loop for model correction , maintenance and monitoring.
• Responsible for Designing entire AI roadmap for project and helping fellow peers to improve their skills and competency in handling production grade machine learning systems.

Cost Optimization

• Developed optimization engine using differential evolution to calculate minimum cost while ensuring all safety checks are performed.
• Devised and build architecture for application and deployed on on-prem server.

Face Mask Detection

• Designed two staged deep learning model consisting of MTCNN and feed forward network to detect whether person is wearing mask or not. Abiding to new rules due to spread of COVID.

Intrusion Detection

• Built Deep Learning based intrusion detection system to detect people in unauthorized areas or unsafe areas using Faster RCNN model with inceptionV2 backend. Integrated system with CCTV cameras using RTSP at construction sites all over India.

Invoice extraction using Deep leaning

• Design deep learning model using Bi-directional LSTM to extract texts of number of key fields from given receipts and save texts for each invoice image, to extract key information from un-structured or semi-structured data.

Auto Learning for ML on cloud

• Built re-trainable machine learning pipeline on Azure.