Jul 2021 – Dec 2022 · 1 yr 5 mos · Bengaluru, Karnataka, India

• Developed search ranking models to rank 1.5 million products on the search results page in our B2B eCommerce
• platform, optimizing for conversions and Gross merchandise value. Serving a million monthly active retailers.
• Built CTR, User cohort personalized CVR prediction models with explore-exploit strategies.
• Deployed real-time services using Docker, Kubernetes, Redis and batch services using SOLR.
• Reduced the feature look up latency P99 from 50ms to 5ms, by caching with Redis and avoiding interregional hops.
• Improved the Search to Add to Cart conversation by 20% and Gross merchandises value by 10%.

May 2019 – Mar 2020 · 10 mos · Guwahati Area, India

• I am leading the Institute’s director highly funded 4i lab’s project RAMAN- The Humanoid Robot.
• I co-ordinate and mentor various other projects like Rubik's cube solver, Robotic arm, Competitions, etc

Mar 2018 – May 2019 · 1 yr 2 mos · Guwahati Area, India

• I worked on the chatbot module of the project RAMAN- The Humanoid Robot.
• I also worked on projects human following trolley, Robotic arm.

May 2019 – Jul 2019 · 2 mos · Nagoya, Aichi, Japan

• I worked on Multi-task learning (MTL Networks) on Semantic Segmentation and Object Detection for estimating Quality of Life(QOL).
• I proposed a State-of-art Network achieving Real-time simultaneous Semantic Segmentation and Object Detection over complex cityscapes dataset.
• During the research, I used to read many works of literature on Semantic Segmentation and Object Detection. I tried to understand the concepts of each research paper and gradually developed a good understanding of CNNs and Deep-Learning. I understood the timeline of CNNs developments over these years.
• In the two months of that internship period, I also completed the implementation of the proposed network in TensorFlow. I implemented Dataset preprocessor for handling both the tasks. I did twice the work it was almost like working for two projects. I trained the Network from scratch(ImageNet weights are used for initializing Encoder). I used several techniques like Differential Learning rates, Freezing a part of the graph, Weight Sharing, Transfer Learning, etc.

Nov 2018 – Dec 2018 · 1 mo · Chennai Area, India

• The ePlane Company was started at the world’s largest combustion center at IIT Madras by Prof. S.R.Chakravarthy to transform the mobility landscape. Currently, building drones for industrial applications.
• I developed an image segmentation model based on AdapNet++ surpasses the benchmarks given in the literature with mIoU score of 89.63 and then integrated it into VLocNet++: Deep Multitask Learning for Semantic Visual Localization and Odometry. Optimized and compressed the model for efficiently deploying in mobile objects like drones
• We have to measure tree girth. First, we tried with the Texture image
• segmentation, and the results were not reliable. So, I came up with an idea to project a laser line on the tree’s trunk and generated the mask for the projected laser line. By using this mask over the depth map, we got very accurate measurements.

Dec 2017 – Aug 2018 · 8 mos · Guwahati Area, India