FlowCV Logo
Education

Information Systems and Programming, Innovation by Design, Design Thinking, Data Science, ML, DL, CV, NLP

Full Tuition Scholarship + Monthly Stipend

Bachelors of Engineering (Hons.)

First Class Honours + Scholarship, 1.47% Acceptance Rate

Undergraduate Thesis @ Harvard-MIT HST, Boston

Technical Skills
Python|PyTorch|Git|C++|Docker|MATLAB|Tensorflow|OpenCV|spaCy|SQL|FlaskCI/CD|Streamlit|FastAPI|AWS SageMaker|JIRA|Pandas|NumPy|Kaggle (Top 1%)
Professional Experience

Senior AI Engineer 2 @ Panasonic

09/2023 – present | Singapore
  • •Lead the AI Project Development from Ideation to Production

    • Robust hyperspectral based plastic classification for recycling

  • •Building a low latency object detection/segmentation model leveraging synthetic data
  • •Building the synthetic data pipeline using image overlays with unpaired image to image translation using CycleGAN
  • •Building YOLO based object tracking pipeline to make the pick-n-place robotic arm more robust
  • •Trained wavelength based tabular data with extensive augmentation and featuring - compared SOTA algorithms
  • •Collaborating with Agency for Science, Technology and Research (A*STAR) and Sembcorp Industries for deployment
    • 11/2020 – 08/2021 | Singapore

    NRIC Verification using OCR to automate the KYC procedure

  • •Built a web app and microservices-based RestAPI in Flask and FastAPI using Keras-OCR & Tesseract (> 80% TP)
  • •Built a custom ID card detection-segmentation model using YOLOv5 and Detectron2 to achieve mAP > 0.9
  • •Optimised Inference time to be < 200ms per image
  • •Computed inference using PyTorch and OpenCV on both CPU and GPU running CUDA 11.2
  • •Dockerised the WebApp and hosted it on internal servers for testing
  • •Operated Bitbucket’s VCS in the CI/CD pipeline. Wrote extensive sanity check routines along with exception handling
  • •Integrated Postman to track the usage of the RestAPI and SwaggerUI to test the API on any browser
  • •Tracked progress on JIRA
  • •Worked on speech-to-text model with speaker diarization using Watson API to analyse customer calls for sentiments

    • Computer vision based plant health and phenotyping monitoring system for indoor farms

  • •Built Plant phenotyping pipeline using YOLOv5 and Detectron2; Achieved mAP > 0.9
  • •Operated in-house data annotation for the POC and collaborated with data annotation agencies for large datasets
  • •Maintained training log and stats using Weights&Biases and Tensorboard
  • •Tested autonomous mapping pipeline using T265 and RGBD by using SIFT for Feature Extraction
  • •Implemented DeepSORT for object tracking and to uniquely count objects in a video
  • •Leveraged different imaging sensors (stereo/ NIR/ multispectral/thermal) to determine plant morphology and quantify different levels of plant health
  • •Integrated logging and unit tests in the development
    • 09/2021 – 10/2021 | Singapore

    Pirate ship object detection/segmentation using both synthetic and real data

  • •Used GCP compute engine to train a object detection/segmentation model leveraging transfer learning
  • •Used both RGB and NIR images and videos for training data
  • •Experimented with Unpaired Image to Image Translation using CycleGAN and Pix2Pix to simulate different surroundings and increase the robustness of the model
    • 12/2018 – 08/2020 | Singapore

    Novel computer vision-based artificial whiskers to detect micro-vibrations underwater for deep ocean object tracking

  • •Deployed the model and evaluated FPS on various edge devices - Raspberry Pi 4, ODROID-XU4 and Jetson Nano
  • •Used Adaptive Thresholding, Sobel-Canny, Feature Matching, Dilation-Erosion coupling, Dense Optical Flow to track the motion of whiskers and contour moments to quantify coordinates in the frame
  • •Utilised the time-series coordinates data to train ARMA and ARIMA for prediction
  • •Used Moving Window Standard Deviation(MWSD) to track minimal temporal changes
  • •Handled missing coordinates before model inference using Facebook's Prophet TS library
  • •Tested the real-time detection model to TFLite to increase the FPS
  • •Corrected Lens distortion using chessboard based camera calibration to calibrate the Camera Intrinsic Matrix
    • 05/2017 – 01/2018 | Boston

  • •Compressed Sensing Based Image Reconstruction in 3D Micro-Bioimaging published in the Royal Society
  • •Mobile-camera based Expansion Mini Microscopy (ExMM)
  • •One of only 2 students from India to get selected for this position
    • Notable Projects
  • •DeepSafe - Open Source - Web based - DeepFake Detector! – GAN, Flask RestAPI, Docker, Google Cloud, Streamlit
  • •Biceps Curl Game - Compete with friends-see who can do more biceps curl! – Pose Estimation, 30 FPS, 32 Key-points
  • •TL;DR WebApp - Summarise research paper abstracts in 1-click! – NLP, Docker, Heroku, AWS EC2, Summarisation
  • •Video Super Resolution - Upscale your videos to 4X higher resolution! – ISR, Artefact Cancellation
    • Academic Publications and Articles

    Academic Publications

    Robust Spectral Feature Extraction and Classification using Hyperspectral Imaging and Unsupervised Deep Learning with Spectral-Preserving Data Augmentation [Ongoing]

    Awards and Media Features

    BITSAA Global 30 Under 30 | Hyperloop India - Finalist, SpaceX Hyperloop Pod Competition

    Granted ~30000 SGD for the design of Advanced Disaster Mitigation Module(ADMM) | Lockheed Martin, USA

    References
    Dr. Ngai Man Cheung, Assistant Professor, SUTD, Singapore
    Dr. Jeffrey Karp, Professor, Harvard University, Boston
    Dr. Y. Shrike Zhang, Assistant Professor, MIT, Boston
    Dr. Pablo Valdivia y Alvarado, Assistant Professor, SUTD, Singapore