Kunal Suri

Triage Analyser-

• Created a solution that predicted whether a vehicle in accident was completely or partially damaged using tabular and image data for one of the largest Auto Insurance Solution Provider in US
• Used XGBoost, DeepFM, embeddings from images, UMAP, HDBscan to build the model that improved the existing Triage process by 20%
• Implemented concept drift detector to identify changes in data distribution that helped in deciding when to retrain the model
• Incorporated Explainable AI in the model using SHAP to explain the predictions to end user
• Developed the model on MLFlow, AWS, Docker, Pytorch, Kubernetes

Vehicle Damage Detector

• Created a solution that identifies car components in the images and localizes the damage in an image for one of the largest Auto Insurance Solution Provider in US ending up reducing turn around time for resolution by 70% in damage estimation
• Used YOLO,UNet,FCN and developed a custom 130 layer deep CNN to build the model that achieved 99.76% F1 Score in identifying 859 different parts in the car from images, ~80% IoU for bounding boxes of damaged parts, and mAP of 70 for segmenting car components.
• Led a team of 40 offshore resources which was responsible for annotating car images with components and their damage types
• Developed a framework of semi-supervised learning to expedite labeling of images by creating an ensemble of 15 CNN models
• Developed the model using MLFlow, AWS, Kubernetes, Pytorch

NLU based Chatbot to automate Customer Service

• Created a solution that automates answering queries of medical sales representatives thereby reducing workload of customer service execs by 60%
• Used cutting edge transformer based language models to build the chatbot that achieved an F1 of 87% and EM of 92% for the answers
• Led a team of 50 onshore and offshore resources to identify questions with same intent and created a model that automated identifying such questions expediting the whole process
• Developed the model using MLFlow, AWS, Kubernetes, FastAPI, Pytorch, HuggingFace, Spacy

• Worked with one of the largest Indian hotel chains in setting up their analytics practice from scratch.
• Used network theory to identify duplicate customers from a customer base of 70 lakhs while ensuring 0% False positive rate. This led to reduction of the entire customer base by 15%
• Re-engineered customer segmentation and created a customer 360 of all customers
• Used a Random Forest/Logistic Regression ensemble to calculate probability of loyalty tier upgrade achieving a precision of 76% leading to 60% increase in revenue
• Used gradient boosted trees to predict customer return propensity and achieved a precision of 80% which led to an increase of 250% in revenue
• Used topic models (LSA) to analyze customer feedbacks to identify factors which led to increase/decrease of ratings leading to increase of half star in ratings for restaurants in Mumbai.

• Worked as part of Analytics and Insights team of TCS on developing solutions for detecting transaction fraud in banking industry
• Used an ensemble of KNN/SVM/Random Forest to calculate probability of fraud on credit card transactions achieving a recall of 85% and precision of 72%

• Worked for one of the largest apparel brands in India with PAN India presence
• Used Gaussian Mixture Models to create customer segmentation by their spendings' behaviour
• Used logistic regression to identify high propensity customers for the End Seasons' Sale. Achieved precision of 70% leading to an increase of 89% in the clients' revenue
• Used Associate Rule Mining to create a product recommendation system to identify cross-sell opportunities