Viji Vennelakanti

Worked with a global team of data scientists, infrastructure engineers, labeling companies, report writers, and researchers to build an innovative computer vision product. I was involved in all aspects of the project and wore multiple hats, including agile technical project manager, product manager, business analyst, data analyst, and QA engineer.

KEY ACCOMPLISHMENTS:

• Worked with the founder to document requirements, business processes, and go-to-market strategies, assisted by data-driven insights.
• Helped select and manage the relationship with labeling companies and the work they performed.
• Planned and implemented computer vision algorithms, training, and testing pipelines.
• Identified project issues and facilitated the team to brainstorm options to resolve them.
• Devised the calculation of accuracy metrics at various stages of the pipeline.
• Wrote the requirements for, managed, and tested an extensive rules-based algorithm to supplement computer vision.
• As the QA analyst, identified automation and outsourcing opportunities.
• Worked with a reporting company to produce complex custom reports.
• Managed the accounts, permissions, storage structure, and billing of AWS accounts.

My client wanted a POC to determine if computer vision using deep learning could identify medical problems from images. I used convolutional neural networks (deep learning) to deliver a successful POC for object detection on medical images with a small sample training dataset of 2,500 annotated images.

The tools used included Google Cloud Platform (GCP), Python, YOLO, and Jupyter Notebook. I set up Ubuntu VM on GCP to use GPU, loaded the necessary software, trained the model using cleansed data, and analyzed the results using object detection metrics. Then I ran another iteration with a larger training dataset and compared results to show improvement.

The client was pleased with the improvements shown and plans to obtain funding to scale the model to accommodate a larger training dataset. I submitted a comprehensive report and recommendations at the end of the POC. The report included an executive summary, approach, data cleansing procedure, GCP setup, training and testing procedure, and object detection metrics.

TRAINING PROCESS:

• Read in over 36,000 images.
• Converted them into tensors.
• Matched them against readings from a CSV file and cleansed the data.
• Ensured that the images were of the correct data type and size.
• Shuffled image and reading data.
• Split into train and cross-validation (CV).
• Test loaded the InceptionV3 model.
• Used its weights on train, CV, and test data for transfer learning.
• Created separate hdf5 files with features, labels, and batch datasets for train, CV, and test batches.
• Trained on a shallow CNN model and transformed the output into the right dimensions for comparison.
• Plotted the loss and per-digit accuracy.
• Calculated the accuracy for training, CV, and test data.

I iterated the above process with multiple hyper-parameters and network sizes along with InceptionResNetV2 to improve the accuracy and used Agile Methodology to keep the project on track and provide visibility. The technologies used to work through this project included Floydhub Cloud Platform, Keras on TensorFlow, Python 3, Jupyter Notebook, and Excel.

This project was implemented in seven countries simultaneously to replace a decades-old, mostly manual business process that had unpredictable cycle times and poor customer service with a simpler, consistent, and mostly automated business process that cut the cycle time to five business days.