Experience

ModiFace

• IT Services and IT Consulting
• 1 - 100 Employee

• AI Researcher

  • Apr 2022 - Present

  - Develop and improve on Computer Vision algorithms used for Beauty/Cosmetic Products at ModiFace - Conduct research and translate to research papers and patents - Develop and improve on Computer Vision algorithms used for Beauty/Cosmetic Products at ModiFace - Conduct research and translate to research papers and patents



Huawei Technologies Canada Co., Ltd.

• Canada
• Telecommunications
• 500 - 600 Employee

• Senior Research Engineer - Computer Vision

  • Sep 2021 - Apr 2022

  • Develop methods to forward the state-of-the-art on video understanding tasks such as video classification, action recognition, audio-visual event localization, temporal action localization, etc.• Develop, port, and fortify cutting-edge research in supervised and self-supervised video understanding research into products via efficient data and model pipelines.• Draft and review research papers and patents pertaining to the developed methods.

• Computer Vision Researcher

  • Sep 2019 - Aug 2021

  • Inducted as a Member of the Patent Review Board. Assessed the commercial value of 10+ AI-based innovations and their potential to become valuable intellectual property in terms of novelty, avoidability, detectability and utility.• Conduct research on different methodologies to leverage multi-modal (audio-visual) mechanisms for video understanding. Design novel architectures (using multi-modal transformers, graph neural nets etc.), loss functions, feature fusion strategies and data augmentation techniques to forward the SoTA in event/action localization and recognition under fully, weakly, and self-supervised settings.• Designed and implemented a video dataset denoising and compression pipeline to speeden and optimize model training on large datasets. Implemented SoTA techniques to reduce model train-to-deploy times by upto 20% and achieved dataset compression rates of upto 66% while inducing a minimal (≈1%) accuracy reduction.• Led a team to design and deploy a hierarchical video tagging system for the Huawei Video platform. The system consisted of image/audio classifiers, object detectors and OCR systems to handle a designated set of tags for video segments. On the field, the system handled increasing tag-set requirements while exceeding the performance goals.• Architected and implemented a semantic image and video search re-ranking system for Huawei Search. • Coordinated a team to develop hand and face detection models used for Rapid Gesture Recognition on Huawei Smart TV. Deployed SSD and YOLOv3 architectures with customized backbone CNNs including MobileNet, ResNe(X)t, FishNet, and SqueezeNe(X)t, to achieve the optimal accuracy-speed tradeoff. Boosted team productivity by streamlining data preparation and visualization, offline novel data augmentation, and model evaluation.

• Research Engineer Intern (Computer Vision)

  • May 2018 - Dec 2018

  • Refined and augmented the Places365 scene recognition dataset. Used model-assisted denoising, human correlation identification, data expansion using Hard Example Mining, categorical merging, and hierarchical labeling for capturing coarse to fine visual concepts. • Developed an end-to-end system for users to learn situation-based English dialogues from English movies for Huawei’s Video Streaming Services. Designed the seamless video segmentation algorithm by fusing shot-based and subtitle-based segments. Situation determination was performed by a scene classifier using a top-down hierarchical voting scheme on the multi-level labels. • Implemented mechanisms to output deployable scene graphs. Designed an attention-based saliency map network to capture salient relationships. A novel IoU based categorical clustering technique produced more unique subject-predicate-object triplets which resulted in richer scene graphs. Drafted 2 new graph evaluation metrics for scoring the structural quality of scene graphs.



Cisco

• United States
• Motor Vehicle Manufacturing
• 1 - 100 Employee

• Software Development Engineer II

  • Oct 2016 - Jul 2017

  1) Nov 2016 - July 2017:a) June 2017 - July 2017:Implemented the user configurable Overhead Compensation feature for the ARAD+ based RSP3. The purpose of this feature was to increase the accuracy of the egress bandwidth when different rewrite values are configured at the ingress and egress by considering the effective pushed/popped dot1q header.b) May 2017 - June 2017:Performed system wide resource analysis (theoretical scaling of all current features on RSP3) for the purpose of increasing Ingress QoS TCAM scale. Implemented the same shortly afterwards and solved various issues which arose due to multidimensional testingc) Nov 2016 - Mar 2017 :Developed the Function Level Heat Map Analysis Tool for SPAG BU (using Python & Couchbase):Implemented the Configurable Linear Convergence (CLC) Formula for Heat Calculation (function level granularity) for the current C based code repository. Designed and implemented the back-end JSON based data model and interfaced the CRUD operations to the DB by parsing and transacting the code base information (such as code commit details, review information, regression timelines etc.) using Python. The tool is used extensively for code refactoring, health monitoring, feature impact analysis etc.

• Software Development Engineer I

  • Aug 2014 - Oct 2016

  1) October 2015 - October 2016:Part of the QoS & ACL Team for the Broadcom ARAD+ based platform; RSP3, in Service Provider Access Group (SPAG) Business Unit. (Programming with C Language)a) July 2016 - August 2016:Delivered the implementation of ingress based IPv6 QoS (DSCP/MPLS Exp based policies). Responsibilities included understanding the current implementation of ingress based QoS, efficient planning to manage resources on the chip (switching between various lookup tables) to deliver the feature, and frequent collaboration with test team to resolve issues at the earliestb) May 2016 - July 2016: Debugged various QoS related issues such as WRED, bandwidth configuration, shaping configuration etc. Developed "show" commands for ACL and QoS to ameliorate on-field debugging.c) December 2015 - April 2016:Developed L3 (IPv4 & IPv6) Access Control Lists on RSP3. This involved meticulous research (with respect to Cisco's implementation of ACLs, resource management and architecture of ingress pipeline on ARAD+ etc) and constructive collaboration with the concerned teams (PI team, multicast team, punt-inject team, Broadcom team etc) to decisively finalise the implementation strategy. It also involved making critical decisions with respect to the behaviour of ACLs on control packets, managing software complexities to handle optimisation of dual stack ACLs and timely resolution of issues by working closely with the test team and Broadcom team.2) August 2014 - September 2015:Part of QoS & ACL Team for ASR 903 and ASR 920 in Service Provider Access Group (SPAG) Business Unit. (Programming with C Language)Developed and debugged features such as traffic classification, prioritisation, marking etc. for the products RSP1 (ASR 903) & RSP2/Striker (ASR 920).