Harindu Ashan

Experience

• 

  Zone24x7

  May 2017 - Dec 2017

  Associate Trainee Electronics Engineer

  • Designed a complete IoT enabled Dressing Occupancy tracker that can communicate to mother device over low-power Wi-Fi signals (Sub-Giga Hertz range) to track down the number of customers entered into/exited from the dressing which can use for statistical purpose. This has improved the efficiency of the store cleaners by more than 67%.

• 

  University of Moratuwa

  Jan 2018 - Aug 2019

  Final Year Student

  • Implemented an Activity Recognition Framework (ARF) using Deep Learning models and traditional feature extraction methods to increase the inference speed by 10%, and to provide the results through a sentence or a paragraph.• Evaluated Deep Learning models empirically to select the best model out of them, to get Video Action Recognition.• Collaborated to design a web GUI for demonstration purpose.

• 

  Zone24x7

  Jan 2019 - May 2021

  • Implemented a Box Detection software with SSD-Lite model for USA-based Production management company that can automatically identify the number of items/parcels issued and returned in regular basis thus the manual counting has been eliminated completely. This is designed on CPU based platform powered by OpenVINO hence the total cost of the project incurred is reduced down by 50%.• Created the custom dataset for the Box Detection software by using CVAT toolbox hence the quality of dataset and the accuracy improvement of the system have been gained more than 30% with mAP of the system 85% with sensitivity of 89%. • Evaluated Background correction system for Passport Photograph, project by Department of Canadian Immigration, to alert the alignment of the photograph to the requirements issued by Canada Immigration Department for Passport applicants while capturing it by using the Machine Learning Algorithms so that the mistakes done by applicants could be reduced by more than 60%.• Developed the Person Detection models using Tiny-Yolo and SSD-Lite on fish-eye camera inputs into the Cognitive Vision Analytic Toolbox to facilitate the platform. Show less • Implemented the hardware services such as Bar code reading, RFID reading, reading IoT tags in the Android powered KIOSK device using Android Studio to collect store information and payment information so that time spent by operator can be reduced down 45%.• Designed Android Apps to enable inter-communication with services thus the services explained above can be used to integrate with main App of the KIOSK device. • Used Android Testing features such as JUnit, Mockito, etc. to evaluate work-ability of the code functions developed therefore the code quality has been improved by greater than 50% with less time spending to implement Apps. Show less

  • Machine Learning Engineer

    Jun 2020 - May 2021

  • Electronics Engineer

    Aug 2019 - May 2020

  • Associate Electronics Engineer

    Jan 2019 - Aug 2019
• 

  ADA

  Jun 2021 - Sept 2021

  Data Scientist

  • Collaborated a project which is the churn prediction in telecommunication network/service, to research different data processing and data cleansing techniques.

• 

  Analog Inference

  Sept 2021 - Jun 2023

  • Developed python scripts for a few tasks such as create custom image classification data sets, transfer weights from one model type to another, quick training codes with PyTorch-Lightning, re-modify models’ backbone/neck/head with other models and train them, etc. to improve operations speed.• Collaborated to evaluate the License Plate Detection Model onto the Hardware Accelerator Processor by using the existed Model repo so that the viability of model being deployed in the hardware could be assessed. • Evaluated YoloV5 model in High-Definition (HD) image processing using the hardware and optimized the model to deploy the model so that the hardware can be fit the heavy weight models too with competitive frames per second metric relative to the rest of the embedded hardware accelerators. • Led a local team of 10 engineers across different projects and reported daily updates to the supervisor in USA. Show less • Changed the Focus Layer in YoloV5 model to remove the slicing operation at input by changing attributes of one convolution layer such as kernel size, stride, padding, etc. which can mimic the same operation as what slice operation and a set of convolutions do thus reducing additional calculation driven by hardware processor.• Introduced an automated process to modify Max-Pool Layers which have kernel size greater than 3 with odd format (5, 7, 13, etc.), to keep kernel size 3 with multiple Max-Pool layers thus enhancing processing speed of the hardware processor by 12%.• Developed a framework to convert model weight files built with PyTorch to Hardware supported structure through the facility from ONNX format so that it has reduced time spent for manual conversion by more than 80%.• Incorporated YoloV5 based framework for other tasks such as audio classification with Mel-Spectrogram features to use the same model architecture for multiple tasks thus the Hardware resources are saved up to 30%. Show less

  • AI Engineer

    Jun 2022 - Jun 2023

  • AI Engineer

    Sept 2021 - May 2022
• 

  TamoTech

  Apr 2022 - Jun 2024

  Artificial Intelligence Engineer (L2)

  • Created Pandas and Seaborn based Visualization techniques for the Audio Data, to visualize distributions of each feature so that understanding the data can be improved.• Incorporated Acoustic Feature processing techniques to extract audio features in patients’ audio clips to inject patient classification problem thus it can raise accuracy by 10% with sensitivity of detecting patient by 4% with relative to the baseline model.• Used various Machine Learning models and Deep Learning models to cross-check the relevant model for patient classification problem. Tree based models such as Random Forest, XGBoost, and CatBoost that are shown the significant improvements of 5% against the Deep Learning models such as ResNet and Simple CNN models.• Embedded Weight and Biases, aka WandB, and TensorBoard to capture the metrics and related information through dashboards so that the model versioning and controlling have been improved. Show less