Harsh Maheshwari

AI Researcher

United States7 yrs 10 mos experience

Most Likely To SwitchHighly Stable

Key Highlights

Fast-Paced Promotion at Amazon
Gold Medal for Best All Rounder at IIT Kharagpur
Highest Technology Award at IIT Kharagpur

Stackforce AI infers this person is a Machine Learning Engineer with expertise in E-commerce and Automotive Technology.

Contact

Skills

Other Skills

Adobe PhotoshopAlgorithmsAmazon Web Services (AWS)Analytical SkillsBig DataBig Data AnalyticsC (Programming Language)C++Data AnalysisData ScienceData StructuresLeadershipManagementMicrosoft OfficeProgramming

About

I am an Applied Scientist at Amazon, working in the Advertising Trust team. Throughout my career I have worked on projects in various domains of machine learning, and I have helped deploy many such projects at a production level. I am a graduate from IIT Kharagpur with a dual degree (B.Tech + M.Tech) in Electrical and a minor in computer science. Since then I have enjoyed applying my knowledge of Machine learning and deep learning to solve unexplored problems. I excel in a collaborative environment, where I can use my teamwork and leadership qualities to drive my work. These abilities and my perseverance have helped me achieve multiple endeavours like - A Fast-Paced Promotion at Amazon - A Gold Medal for Best All Rounder at IIT Kharagpur by Education Minister of India - The Highest Technology Award at IIT Kharagpur

Experience

Amazon

3 roles

Applied Scientist 2

Aug 2024 – Present · 1 yr 7 mos

Working in International Seller Growth Team to work on Project Amelia and GenAI driven projects to enable seller growth

Applied Scientist 2

Mar 2024 – Sep 2024 · 6 mos

Worked in geospatial domain. Developed a model to determine if the given address is complete or not using neighbouring address.The model outperformed the normal classification model by 23% and validated the effectiveness of model across various geographies
Worked on predicting geo codes for a given textual address using existing database of address-geocode pairs. This is used to get geocode for a new address in the system

Applied Scientist 2

Dec 2022 – Mar 2024 · 1 yr 3 mos

1) Conceptualized a deep learning-driven auto-correction module for rectifying text-based imperfection in video ads. Orchestrated a model pipeline encompassing of video inpainting, font type detection, endpoint detection & more
2) Developed Retrieval Augmented Generation(RAG) model using LLM for moderation policy understanding.
3) Fine tuned the OpenLLaMa 7B model through knowledge distillation, leveraging instruction reasoning from a larger model to automate the moderation workflow. Achieved recall ≥ 95% and precision ≥ 90% for multiple defects
4) Worked on various LLM inference library like vllm, TGI etc to optimize model through put and latency
5) Working on task assignment optimisation problem at scale of 100k tasks & 5k solver. Developed simulation using live data, enabling exploration of various strategies & providing extensive data visualizations for decision making
6) Leveraged Contrastive Learning to boost performance by 0.8% over DiffCSE on internal textual dataset

Gan studio

Machine Learning Intern

Jan 2021 – Apr 2021 · 3 mos · India

Worked in a fast-paced development environment on the following
Worked on vocoder (converting Mel spectrograms to audio) using WavGlow, WavRNN, Mel GAN and Multi-Band Mel Gan vocoder. Trained vocoder for multi speaker LibriTTS dataset, outperforming many available vocoders
Worked on Text to speech synthesis for multi-speaker setting with the aim to improve voice cloning abilities
Worked on voice cloning technique and implemented state of the art model like Ge2e, SASN, Resnet34L, Resnet34V2 and was able to clone voice in real time application
Worked on distributed GPU training, Mixed precision training (FP16)
Implemented and imported the entire code base to PyTorch Lightning framework for the flexibility of multi-GPU and TPU training
Implemented batch samplers to optimize the GPU consumption during training for variable-length input data
Implemented Emotion classification model
The overall experience gave me a good experience in handling a large codebase, collaborating with multiple people on same codebase, good coding skills, and an enormous span of knowledge in the field of audio.

Indian institute of technology, kharagpur

3 roles

Teaching Assistant

Jul 2020 – Apr 2021 · 9 mos

Worked under Professor Alok Kanti Deb and Professor Debaprasad Kastha of Electrical engineering department to conduct classes of Electrical Technology course for first year undergraduate students.
Helped in making assignment and correcting copies of students.

Master Thesis

Jun 2020 – Apr 2021 · 10 mos

Worked on medical imaging segmentation problem
Proposed a computationally efficient solution for segmentation of 3D volume
One paper got accepted at Medical Imaging with Deep Learning 2021
Also worked on ultrasound simulation. Specifically analysing the effect of change in frequency with ultrasound simulation. (Paper submitted in Journal)

Bachelor Thesis

Jul 2019 – Apr 2020 · 9 mos · Kharagpur

◦ Worked on video inpainting, with aim to generate temporally and spatially consistent, inpainted outputs.
◦ Implemented many state of the art methods such as MMD-GAN, WGAN, WGAN-GP to stablise GAN training.
◦ Implemented various parameter efficient temporal attention module such as TGM, TSM etc and implemented 3D
temporal PatchGAN discriminator to handle free form inpainting. Secured 0.7767 FID score on FVI dataset
◦ Secured Ex Grade in the final evaluation

Amazon

Applied Scientist Intern

Apr 2020 – Jul 2020 · 3 mos · Bengaluru, Karnataka, India

Worked on noisy labelled image classification
Successful implemented and tested multiple baselines published in top tier conference.
Proposed a method which combines both noise-robust and semi-supervised approach for noisy labelled classification.
Verified proposed method on Synthetically generated MNIST noisy dataset, Amazon dataset and Research dataset (Clothing 1M and Food101N).
Our work accepted in a workshop of AMLC (Amazon Machine learning conference).
Familiar with different AWS services like Sage maker, Athena, S3, AWS Rekognition etc.

Carnegie mellon university

Summer Intern

May 2019 – Jul 2019 · 2 mos · Pittsburgh, Pennsylvania

Worked on three different projects
Achieved State-of-the-art results in domain adaptive object detection. Secured 3.5mAP higher as compared to current SOTA. Link to paper is attached.
Our work accpeted in IJCV journal(Impact factor - 6.071). Paper attached.
Developed an end to end industrial product for bottle recycling plant
Implemented state of the art methods like look closer to see better, Bilinear CNN, Auto augment for fine-grain classification.

National digital library of india

Summer Intern

May 2018 – Jun 2018 · 1 mo · IIT Kharagpur

Summer intern at NDLI. Devised an algorithm for crawler which is able to crawl data from multiple website, hosted by same data base management.

Autonomous ground vehicle research group

Group Member

Feb 2017 – May 2021 · 4 yrs 3 mos · Kharagpur Area, India

We are a group of undergraduate and graduate students who are working to build India's first autonomous self-driving car. Team AGV is one of the finalists of the Mahindra rise prize competition. We have secured the second rank at Intelligent ground vehicle competition held at Detroit, Michigan, USA. My major work in the group is developing software stacks such as solving the problem of pattern recognition, traffic light detection, pedestrian detection using computer vision and deep learning. Also, I have worked on path planning and localisation. I have successfully published two papers here.