Ashish B. Email & Phone Number

Edmonton, AB, CA

Canada

• Designed and developed financial domain-expert GPT platform using fine-tuned Llama 3.1 70B.
• Developed FinGPT to generate benchmark financials, forecast valuations, finance summarizations, short & long financial QnA pairs, calculating financial metrics and values such as PBT, PAT, EBIDTA, DCF.
• Generated synthetic instruction-following dataset, summarisation dataset, and QnA pair dataset for custom-based fine-tuning LLMs.
• Extracted complex data from tables, graphs, and images for text embeddings to fine-tune LLMs using table transformers and papermage.
• Fine-tuned both Vanilla RAG and GraphRAG models, for enhancing contextual understanding for more accurate and relevant generative outputs.
• Fine-tuned both open-source and paid LLMs for finance, real estate, and insurance domains

Canada

• 1. Client Engagement:
• Collaborated with prestigious clients, including Greater Toronto Airport Authorities and Alberta Cancer Institute, to address their unique challenges and contribute to their technological advancements.2. 3D Video.
• Applied expertise in 3D video processing and image segmentation models.
• Developed and implemented advanced algorithms for accurate analysis and extraction of meaningful information from complex visual data.3. Synthetic Image Generation:
• Utilized advanced image processing techniques to generate synthetic images for model training.
• Implemented cutting-edge methods to enhance the quality and diversity of synthetic data, contributing to the robustness of machine learning models.4. Application-Specific Synthetic Data:

Ottawa, Ontario, Canada

• Project 1: Home Cage Analytics
• Implemented an analytics system for mice inside a cage by extracting and processing coordinates and velocity data.
• Utilized Deep Lab Cut software and a ResNet101 trained model for accurate tracking and analysis.
• Automated neuroscientists' workflow, reducing manual work by an impressive 95%.Project 2: Rope Quantifying
• Developed an end-to-end solution using a fine-tuned ResNet101 algorithm to estimate the amount of string pulled by mice.
• Achieved an outstanding accuracy of 99%, surpassing manual scoring by 70%.

• 1. Automatic Visual Tagging for Fashion Attributes:
• Designed and launched an end-to-end automatic visual tagging system for 350 fashion attributes in images.
• Achieved an impressive accuracy of 97% by fine-tuning FasterRCNN and EfficientDet models.
• Anticipated a substantial impact on fashion e-commerce, projecting a 30% increase in business.2. Synthetic Image Generation:
• Generated 1 million synthetic images using a combination of MobileNet and Segmentation techniques for training the model.
• Enhanced the diversity and quantity of the dataset, contributing to the robustness and generalization of the machine learning model.3. Visual Search Engine & Image Retrieval System:

Virtual

Mentored the team on the solution easy navigation and assist for Electric Vehicles given by VMWare India Pvt Ltd. An android app for a user that navigates him/her to the nearest charging station provided with features like battery swap & battery delivery system. With USP as business use cases, we built a web application as a dashboard for the grid owners.

Hyderabad, Telangana, India

- Worked on developing a vision-based cricket analysis product that extracts key parameters from the batsmen, bowler, and fielder. - The whole platform was designed to deploy on the Nvidia Jetson Nano with a 60 FPS camera.

• 1. Real-Time Object Detection and OCR Deployments:
• Pioneered the development and deployment of 30 real-time object detection and 10 OCR products.
• Successfully deployed on both local GPUs and cloud GPUs with parallel GPU processing.
• Resulted in a significant business impact, increasing client revenues by 40%.2. Model Optimization for Edge Devices:
• Transformed heavy-weight models into lightweight models specifically tailored for Nvidia Jetson Nano and Raspberry Pi deployments.
• Achieved a remarkable 50% reduction in computational costs, optimizing performance for standalone products.3. Custom Dataset Generation:

Problem Statement provided by Samsung R&D: Designed and implemented Heart Rate estimation using a Mobile Camera using Computer Vision algorithm. Defined with controlled environment scene of estimation, designed the algorithm in MATLAB and Python. Implemented PoC with Python by deploying it in Heroku and visualized output in an Android phone.

IIIT Hyderabad

Developed and implemented ConvNets and ResNets - 32, 50, 100, 150 for the automatic detection and grading of diabetic retinopathy with 90% accuracy

Master Of Engineering - Meng, Electrical & Computer Engineering Specialization In Applied Artificial Intelligence

Advanced Certification And Foundations Of Artificial Intelligence, Ml, Dl, Computer Vision, Nlp

Bachelor Of Technology, Electronics And Communications Engineering