Christen Millerdurai

Experience

• 

  Amazon

  Jan 2018 - May 2018

  Application Engineer Intern

  Responsible for developing a statistical model to analyse tickets reported by both customers and services, to provide insights to reduce the number of tickets generated, by text mining, NLP, scrapping.

• 

  Signzy

  May 2018 - Dec 2020

  Highlights:==> E2E Development of Signzy's in-house multi-orientation Natural Scene Text Recognition pipeline for extraction of text from ID cards, Documents, payslips, etc in PDFs and images. The pipeline was first RnDed using python and then ported to a highly scalable in-house batched C++ implementation, to serve 4QPS in a T4 instance. The accuracy is as good as the state of the art solutions, so we replaced third-party OCR with our on-premise solution in prod to reduce incurred costs. (~1-year effort)==> Developed an Auto-ML platform [Ai, Backend, some fronted code] exclusively for B2B customers, so they can annotate, train and deploy AI models without our help. Our USP was NER tuned to Id/document extraction scalable to any number of different cards throughout the world. Feel free to contact Signzy sales, if you want to know more. (~1-year effort)==> Developed an E2E face authentication and face search system for enterprise, where the search is scalable to ~1M users (as far as we have tested). The main challenge was processing the face encodings to search many users, with high precision and response time. We did a lot of tweaking and in the end were happy with results.==> Developed and implemented a highly scalable GPU batching interface for Python and C++ common to all APIs we use, to reduce development time for each new API, and also increase the throughput for compute intense APIs. ==> E2E Development of MRZ extraction solution to automatically detect and extract MRZ data from natural scene images (mobile photos, scans, xeroxes, etc). Benchmarks show we are one of the best!.==> Did research on text detection and developed a novel text detection NN, published a white paper about it, https://tinyurl.com/yxkwwoum . Code in my GitHub.Other than this, worked in id-extraction, id-classification, id-cropping, document forgery(copy move detection), liveliness detection, mobile id-extraction solution(offline), etc. Show less ==> Implemented a CNN based classifier for identity cards.i.e, Aadhar, DL, PAN, Passport for resource constraint mobile architecture [ Squeezenet ] and serve it using tensorflow mobile.==> Implemented a spoof detection mechanism for detecting fake videos,i.e videos where the person's face is a consequence of identity theft.==> Working on a resource constraint document segmentation algorithm, i.e extracting foreground objects like identity cards, papers, etc, from the background using deep learning, to make the Cx experience as fluid as possible.==> Working on an E2E Liveliness detection model API where the request to response time should be under sub 5 seconds i.e real-time, where the model is an ensemble of face recognition, gesture recognition, smile detection and face sequence detection. Show less

  • Machine Learning Engineer - II

    Oct 2019 - Dec 2020

  • Core A.I SDE

    Oct 2018 - Sept 2019

  • Machine Learning And Computer Vision Intern

    May 2018 - Sept 2018
• 

  Max Planck Institute for Informatics

  Jul 2021 - Feb 2024

  Real-time Egocentric Pose Estimation Using Event Cameras D6. Visual Computing and Artificial Intelligence==> Denoising point clouds using spectral analysis and sampling patterns.==> Multi-view Mesh parameterization of hands.==> Monocular interacting hands mesh parameterization using event cameras.

  • Master Thesis Student

    Dec 2022 - Feb 2024

  • Research Assistant

    Jul 2021 - Jan 2023
• 

  Starryai

  Jun 2022 - Aug 2022

  Machine Learning Engineer

  Worked on Denoising Diffusion models to generate artistic images.

• 

  Freelance

  Oct 2022 - now

  Computer Vision and Computer Graphics Freelancer
• 

  FZI Forschungszentrum Informatik

  Jan 2023 - Mar 2023

  Research Assistant

  Regressing the corner points (Dimensions) of delivery boxes (Parcels) in a conveyor belt. The algorithm had to handle deformed and undamaged parcels passing through a conveyor belt. The input modality were point clouds.

• 

  AUDI AG

  Apr 2023 - Sept 2023

  Machine Learning Intern

  => Integrating Diffusion models to improve the workflow of designers.=> Created a system to identify and remove copyrighted images that diffusion models might have generated (avoid designers from sourcing design ideas from copyrighted images)=>Created a pipeline for training and finetuning custom diffusion models with minimal input from the user.

• 

  FZI Forschungszentrum Informatik

  Sept 2023 - Dec 2023

  Research Assistant

  OCR for coding videos.

• 

  Deutsches Forschungszentrum für Künstliche Intelligenz (DFKI)

  Feb 2024 - now

  Researcher Phd Candidate