My skills
Programming Languages
Python
C++
Artificial Intelligance
Machine Learning
Deep Learning
Computer Vision
Speech
Natural Language Processing
AI Tools
Pytorch
Tensorflow
Keras
Jax
TensorRT
ONNX
OpenVINO
Caffe
Dev tools
Docker
Triton Inference Server
Git
AWS
Azure
GCP
My Experience
Face AI Research
Lead Machine Learning Engineer, HyperVerge Inc., January 2022 to Present
Machine Learning Engineer-II, HyperVerge Inc., January 2021 to Present
AI Engineer, HyperVerge Inc., June 2020 to January 2021
-
Currently heading the Face AI Research team building face-based identity verification & fraud detection algorithms, Fair Face Recognition algorithms, 2D & 3D face synthesis models, and efficient deep learning architectures for FR.
-
Built the best face recognition algorithms in the nation and some of the best in the world, ranking among the top 5 globally on the prestigious US NIST's Face Recognition Vendor Test (FRVT) 1:1 and 1:N leaderboards.
-
Leading the new AI research initiatives and go-to-market in the organization like deepfake detection, visual speech recognition, and face-based remote health assessment.
-
Worked on approximate nearest neighbor (ANN) algorithms for efficient billion-scale face search and de-duplication systems. Worked extensively with ANN libraries like Milvus and Faiss.
-
Built highly accurate rotation invariant face detector and deployed it in production, which more than 100 clients in different countries use, detecting millions of faces every month.
-
Developed efficient distributed deep learning training pipelines for multi-node and multi-GPU training workflows with non-trivial deep learning architectures.
-
Took deep learning optimization and deployment tools like TensorRT, Triton Inference Server, ONNX and OpenVINO mainstream throughout the organization.
-
Responsible for setting up all the QC pipelines and documentation for the face recognition pod. Conducted numerous POCs with potential clients to assess value addition and understand new product directions.
-
Responsible for automating POC and digitization pipelines which decreased the manual efforts from days to a few hours.
-
• Managed and mentored more than 10 interns and ML Engineers to help them successfully achieve their goals.
-
Worked closely with the engineering team to take our models to production. Also worked with product and business teams for product market research, ideation, and understanding client requirements.
Document Fraud Detection using Deep Convolutional Networks
Deep Learning Intern, HyperVerge Inc., May 2019 to July 2019
Hyperverge is a team of young and amazing people founded by graduates from IIT Madras, funded by investors from silicon valley like NEA, Milliways Ventures and Naya Ventures. Hyperverge develops AI systems for FinTech and GeoSpatial analysis. Some of the companies using services from Hyperverge are Jio, Vodaphone, FE Credit, Airbus, etc. During my time at Hyperverge, I have
-
Developed an ID card liveness detection module used for e-KYC and on-boarding for a Vietnamese credit company.
-
Experimented with different Deep Convolutional Network architectures like VGG, ResNets, ResNeXts, DenseNets and achieved an accuracy of about 97% in liveness detection and deployed the model into the production pipeline.
-
Experimented with the latest deep learning training strategies like Multi-Task learning and Deeply Supervised Learning to make the model generalize on any kind of ID card around the world.
Person Re-Identification for Security Surveillance systems.
Deep Learning Intern, Uncanny Vision, May 2018 to July 2018
Uncanny Vision is a startup based in Bangalore funded by companies like Renesas, Qualcomm. Major work done here is on smart surveillance i.e, the application of Deep Learning to Security Surveillance systems and has many patents in the area. Uncanny Vision has contracts with Japan Government, Renesas, Bosch, with Bangalore metro and many other companies. As a summer intern,
-
Kickstarted the research on Person Re-Identification at the company for smart security surveillance.
-
In particular, I worked on the problem of Person or Object Re-Identification using pose estimation. In just 2 months, I have built a Person Re-Id algorithm which is ready to incorporate on the security camera module.
-
Built a deep network with a Siamese like architecture which uses pose estimation along with a special similarity estimation layer which shot the accuracy near to the state of the art at 85% on a very challenging real-life data from a partner company.
Semantic segmentation, localization and classification of of skin cancers using Deep Leraning.
Deep Learning Intern, Chainrule.ai, March 2018 to May 2018
Chainrule.ai is an early stage startup based in Silicon Valley, San Francisco Bay Area. The main goal of the company is to create a user end applications suit for doctors to help them detect various types of cancers in different parts of the human body effectively in less amount of time. Work is being done on detection of different types of cancers like skin cancer, liver cancer, lung cancer, breast cancer, intestine cancers, etc. As an intern,
-
Worked on building deep learning algorithms for segmentation and classification of different types of skin cancers.
-
After experimenting with different types of convolution-deconvolution networks like U-net. I have achieved Jaccard index of about 0.7 using a U-net like model in segmentation.
-
I have used an ensemble of multiple ResNets and have achieved an accuracy of about 92% in classification.