I'm Ashwin, a Machine Learning Engineer, based out of Bengaluru, India.

About Me

About Me

Thanks for visiting my profile. Having 5 years of overall experience, I currently work as an AI Engineer solving Computer Vision and NLP problems in Automotive, Retail and Supply Chain Industries. My skillset includes expertise in programming languages such as Python and experience with popular deep learning libraries such as PyTorch and Tensorflow. I also have a strong understanding of MLOps with Kubeflow and mlflow, which allows me to effectively design, build, and maintain scalable, highly available, and robust AI tools.

In addition, I also experiment with the Deep learning models on Edge devices like Nvidia Jetson/Google Tensor for on-device inferencing. Apart from AI, I am an advocate of Digital privacy and self-hosting. I have hosted my own NextCloud, Password Manager, Code Copilot and even Llama chat running on my own Hardware (Jetson).



Mandatory Links ;)

   

Work Projects

C-V2X Safer Road Intersection:

Been a Part of C-V2X Team for building a PoC on Safer Road Intersection for Vulnerable Road User with real-time Video Analytics solution altering all ROI objects in the vicinity using Nvidia Deepstream.

My Role:

  • Prepared object detection/segmentation CCTV Survelliance Video Datasets (Synthetic and Real)
  • Finetuned Yolov6 with the cctv dataset with AWS Sagemaker Multi-Gpu training Jobs.
  • Quantization with TensorRT for efficient inference on GPU device within 1.5ms/frame (BS =1) and deployed with Triton Inference Server.
  • Model Training and Testing of Vision Model with Synthetic Data generated by Nvidia Omniverse and Carla, for tuning the model to detect objects with occlusion and artifacts for 2D object detection / semantic / instance segmentation.

Tools and Tech:

Python, C++, Nvidia Deepstream, Triton Inference Server, TensorRT, Kubeflow, Docker, Kubernetes

Machine Translation

Led the development of a Neural Machine Translation system for translating various Technical documents,Part attributes and Descriptions from English to Canadian French and Mexican Spanish which saves US $0.15 per word translation cost to the company.

My Role:

  • Prepared the dataset by scraping existing human translated docs using various Python text processing toolkits like NLTK, Spacy.
  • Finetuned Helsinki MT Eng-Spa model to obtain BLEU score of ~34 using Hugging face transformers.

Side/Hobby Projects:

Automotive Domain LLM

Developed a PoC for autocompletion of sentences in Automotive Maintenance Documents using Text generation with GPT2.

My Role:

  • Developed Training and test dataset of 200k sentences and 40k sentences respectively from 100+ Mechanical/Automotive Engineering pdf books.
  • Using Hugging face Transformers Fine-tuned the GPT2 model to produce high quality automotive domain text predictions.
Model Link

Blogs:

Kubeflow pipelines and SageMaker ML workflow

This Blog shows how to build your first Kubeflow pipeline with Amazon Sagemaker components using the Kubeflow Pipelines SDK.

Blog Link

Mlflow and Nvidia Triton

This Blog explains about Serving and Managing ML models with mlfow and triton inference server.

Blog Link