We conducted an extensive evaluation of small language models for their effectiveness in extracting entities and relations to generate knowledge graphs from text corpora. Our assessment included models like Phi-2/3 by Microsoft and Gemma by Google, with a particular focus on various prompting techniques. The primary objective was to compare their performance against leading models such as LLaMA2 and Mistral, as well as large language models (LLMs) like GPT-4.

Automated claims assessment using AI for automotive and travel insurance industry

Car Insurance - Processed ~5 Million images for multiple deep learning tasks

1. Trained, tested and integrated an ensemble of instance segmentation (MaskRCNN), semantic segmentation (UNET) and object detection (SSD) models, for damages (>90% accuracy) and part detections (>97% accuracy)

2. Trained classification models in Keras with focal loss to handle unbalanced dataset, inorder to classify car model (>95% accuracy), color (>90% accuracy) and damaged parts (>80% accuracy)

3. Deployed 15 deep learning models on TensorFlow serving for efficient inference and dockerized it for scaling

4. Engineered the Flask back-end of the product and integrated it with serving and react using http requests

5. Solely generated a codebase of 100 scripts and 20,000 lines of code from scratch

6. Integrated AWS optical character recognition APIs to parse VIN number, registration and license plates

7. Developed YOLO object detection model for odometer reading in seven-segmentdigitformat

8. Incorporated databases like Redis, AWS S3 and AWS Dynamodb for data storage and retrieval on the fly

9. Configured Supervisor, Gunicorn, Redis Queue and Nginx to manage processes during production

Travel Insurance

1. Achieved a 100% accuracy on claim settlement amount without training any deep learning model

2. Extracted text from customer tickets, refunds and bank documents using AWS OCR apis

3. Compiled customer, travel and payment information from each document using text mining techniques

Provided API as a Service for customers in FMCG, Finance and Oil & Gas Industries

1. Trained object detection models to solve customer requests like logo detection, invoice digitization, implants

detection, person counter, text extraction, pipeline leakage detection and rust detection

2. Programmed the business domain workflow in flask rest-apis and hosted them on AWS and GCP instances

‘GynaeCam’ - A low-cost cervical cancer screening device with ML aided clinical workflow

1. Built object detection (80 mAP) and image classification (75% accuracy) models using TensorFlow and Keras APIs, to localize cervix and classify it into stage 1 versus stage 2 of cervical pre-cancer respectively

2. Optimized the models using TensorFlow Lite for integration in the android application

3. Studied 30+ research papers to get more insights of the problem space and pre-processing approaches

4. Led a team of five, to build a full-stack integration with deployment of AI

Built a full-stack integration for deployment of AI, along with image classification and object detection models in ipython on Keras and Tensorflow. Implemented both of the above models in an android app through Tensorflow Lite. Studied 30+ research papers to deep dive and get more insights of the problem space. Frequently field tested the hardware device and the android app for feasibility

Uterine cervix image segmentation to detect pre-cancer - ML/AI - Computer Vision - Deep Learning - Tensorflow 2.0 - Python - OpenCV - CUDA