Nikita Surya
Development
Hyderabad, India
Skills
Machine Learning (ML)
About
NIKITA SURYA's skills align with IT R&D Professionals (Information and Communication Technology). NIKITA SURYA appears to be a low-to-mid level candidate, with 4 years of experience.
View more
Work Experience
Applied Scientist
Perfios Software Solutions
March 2021 - April 2022
- Headed in designing and streamlining digitization of loan approval documents for a leading bank in India. • Digitized 10M documents which created a potential revenue of $2M revenue for the company. • Integrated data drift detection and misclassification analysis at large scale. • Responsible for building a STP journey for fast processing of few documents. Worked on compressing deep learning-based models using different techniques like ONNX, static and dynamic compression.
Student Researcher
Cybernetics Lab
September 2019 - December 2020
- Increasing the quality of rail transport through image data-based damage pattern recognition in rail vehicles- QUISS Discontinuity detection, OCR of wagon ID, rescale wagon images, discover common objects in a wagon type, estimate the performance based on this information
Cybernetics Lab
March 2020 - October 2020
- Developed an automatic approach for global interpretability of image-based classifications Applied feature visualization techniques to extract meaningful concepts responsible for neural network prediction. Using TCAV to get importance scores of the concept images for a particular class of images
Applied Scientist 2
Perfios Software Solutions
April 2022 - Present
- Explainability (XAI) of CNN based models. • Publication: Developed a method called ECLAD to interpret image-based classification models using global explanations and a quantitative score. Co-authored and published this research manuscript ECLAD: Extracting Concepts with Local Aggregated Descriptors in the prestigious Pattern Recognition journal. • Collaboration: Spearheading the collaboration between RWTH Aachen, Germany and Perfios to research explainability in object detection models. Leading a team of 3 Applied Scientists and 2 Software Developers in this collaboration. • Operationalised explainability method for classification models capturing both local and global explanations. This method aids developers by giving explanations in the case of data drift in production and detecting bias in data. Leading a team to build models for generic Key-Value prediction in scanned documents using Layout LM backbone. Streamlining the process to Group, Label, and Link the entities. This will be used by financial institutions as a service for redacting personally identifiable information (PII) and minimize the workload by ~60% by the data processing team. Built a NER model that process data for recommendation models by extracting entities from UPI, NEFT, and RTGS narrations for top 20 banks in India. Achieved 12% better performance than baseline methods and generated a revenue of $1M.
Education
RWTH Aachen University
Master's in Robotics Systems Engineering
October 2018 - December 2020
