SAJAN R. GOHIL

B.Tech in Information and Communication Technology - Pandit Deendayal Energy University (formerly Pandit Deendayal Petroleum University) (Graduated: 2020, 8.94 GPA)

• Gohil, S., Lad, A. (2022). Kidney and Kidney Tumor Segmentation Using Spatial and Channel Attention Enhanced U‑Net. In: Heller, N., Isensee, F., Trofimova, D., Tejpaul, R., Papanikolopoulos, N., Weight, C. (eds) Kidney and Kidney Tumor Segmentation. KiTS 2021. Lecture Notes in Computer Science, vol 13168. Springer, Cham. https://doi.org/10.1007/978‑3‑030‑98385‑7_20. Also in: Kidney and Kidney Tu‑mor Segmentation: MICCAI 2021 Challenge, KiTS 2021, Held in Conjunction with MICCAI 2021, Strasbourg, France, September 27, 2021, Proceedings (pp. 151–157). Springer‑Verlag.

Machine Learning Engineer - Glib (Genesis Artificial Intelligence Pvt. Ltd.) (July 2020 - Present)

• Successfully designed and built systems for data extraction from financial documents and helped upgrade the extraction algorithms from rule based to AI first algorithms with more then 95% extraction rate.
• Developed novel as well as existing state of the art computer vision, natural language processing and multi‑modal deep learning architectures for tasks such as named entity recognition, object detection, clustering, image processing etc for document layout analysis.
• Deployed extraction systems on private and on‑premise cloud servers.
• Led a team of 2‑3 people for smaller projects in the same domain.

Deep Learning Intern - Geeky Bee AI Pvt. Ltd. (Jan 2020 - June 2020)

• Built modules for a real‑time weapon detection system using YOLOv3 object detector.
• Designed and developed desktop UI and deep learning inference backend of a face recognition based bio‑metric attendance system using few‑shot learning methods, optimized for cpu inference using Intel’s OpenVino toolkit with performance of more than 20 frames per second.

Software Development Intern - SilCore Technlogy Pvt. Ltd. (June 2019 ‑ July 2019)

• Developed software for controlling a custom programmable logic controller using EtherCAT protocol and storing the data received in a PostgreSQL database.

• Programming/scripting: Python, C/C++, Javascript, bash
• Libraries and frameworks: Pytorch, Transformers, Tensorflow, Keras, PyTorch Geometric, torchIO, OpenCV, Numpy, Pandas, sklearn, etc
• Database management: MongoDB, SQL, SQLite, MySQL, PostgreSQL, Firebase
• Deployment: Intel OpenVino toolkit, Flask, FastAPI, Docker, Amazon Web Services, Google Cloud Platform

Large scale dataset and baseline for Table Structure Recognition (Ongoing research project)

• Problem Statement: Generate a large dataset pdfs containing tables and word/cell annotations from wikipedia and synthetically add styles to tables. Create a baseline algorithm to cluster words in cells and find relation between cells to identify table structure.
• Created a dataset of 1.6 million tables from various wikipedia articles with styles generated by varying different attributes of tables and text.
• Performed word clustering to obtain individual cells with accuracy of 75% tables with perfect cell separation on subset of the data using graph attention network and spatial features.

Facial recognition using optimized neural networks to censor bystanders’ faces using OpenVino toolkit for cpu inference (Development project)

• Problem Statement: Application to detect, recognize and blur unrecognized faces from video footage using convolutional neural networks optimized for cpu.
• Optimized mobile net single shot detector network using quantization and pruning with Openvino toolkit for face detection.
• Used and optimized convolutional neural network for few‑shot face recognition to collect embeddings of known faces.
• Developed inference backend that registers known faces and outputs video with other detected faces blurred.

Parallelized Approach To Identify A Person In Surveillance Footage (Research project)

• Problem Statement: Retrieve person matching the given soft‑biometric description from a surveillance footage.
• Used Masked‑RCNN to perform person detection.
• Trained convolutional networks for different bio‑metric features on extracted person images from different datasets.
• Achieved 88% accuracy by creating an ensemble of the trained convolutional neural networks and used gradient boosting tree classifier to perform fusion of neural network embeddings.

• IBM Data Science Professional certificate , Coursera
• Blockchain Specialization , Coursera
• AI for Medicine Specialization , Coursera
• Deep Learning Nanodegree , Udacity
• Intel Edge AI for IoT Developers Nanodegree , Udacity
• Deep Learning Specialization , Coursera

• Facebook Developer Circle scholarship for Udacity’s Deep Learning Nanodegree
• Intel Edge AI scholarship for Udacity’s Intel® Edge AI for IoT Developers Nanodegree