Swapnil Bhirud

Generative AI : Natural Language to SQL Query

• Developed a Natural Language → SQL query system leveraging Meta’s LLaMA-3 (70B), enabling non-technical stakeholders to retrieve data seamlessly.
• Engineered prompt templates incorporating database schema context, significantly enhancing accuracy and relevance of generated SQL queries.
• Implemented Retrieval-Augmented Generation (RAG) for dynamic prompt generation, significantly improving query accuracy
• Built robust query execution pipeline with validation, syntax checks, and error handling, ensuring safe execution on live databases.

Sales Lead Identification

• Automated web scraping pipeline using Selenium in Python to collect headlines from news articles for potential lead generation.
• Applied NLP techniques by converting headlines into embeddings using Sentence Transformers.
• Trained and deployed XGBoost classifier to classify and prioritize potential sales leads, improving identification efficiency from 56% to 96%.

Demand Forecasting (ongoing)

• Building global demand forecasting models for a leading FMCG giant, covering multiple geographies and product categories.
• Applying state-of-the-art time series and machine learning techniques (ARIMA, XGBoost) to capture seasonality, regional trends, and external drivers.
• Designing scalable forecasting pipeline with automated data ingestion, preprocessing, and model retraining to support real-time decision-making.

AI Motor Damage Localization and Severity Detection

• Created labeled image dataset of car damages using Amazon SageMaker for supervised training.
• Developed a Mask R-CNN instance segmentation model in TensorFlow/Keras (Python) to accurately localize vehicle damages.
• Classified damages into four categories – scratches, minor dent, heavy dent, and tear – achieving 96% accuracy for scratches, and up to 76% for dents and tears.
• Improved fraud detection and reduced human error in vehicle inspection by automating damage assessment, increasing transparency and trust in the claims process.

• Shift from Manual solutions approach to Algorithmic solutions approach : Designed and developed Cab pooling & Fleet optimization Routing Algorithms in python using geometry and graph based approach.
• The algorithmic approach (Djikstra's algorithm) enhanced accuracy in terms of Vehicle Utilization, Occupancy optimization, Idle time minimization, Dead Run Minimization and hence Cost optimization.
• Enhanced margin improvement of the past deals using the built in routing algorithm by around 1.5% amounting to around Rs. 2 Cr.
• Responsible for very competitive proposals, hence improving the hit ratio (client conversion ratio) from 7% to 10%
• EV allocation : Proposed and successfully deployed over 100 EVs for various clients as part of the sustainability and green energy initiative. Also proposed suitable number & geo locations for setting up the charging stations.

Movie Recommendation Engine

• Developed a movie recommendation engine in R leveraging collaborative filtering techniques.
• Implemented two approaches: Cosine Similarity & Pearson Correlation and Non-Negative Matrix Factorization (NMF) to predict user preferences.
• Achieved a hit ratio of 8/10, indicating strong alignment between predicted and actual movies watched by users.

Cab Fleet & Cab Pooling Optimization

• Optimized cab fleet operations by reducing dead runs and idle time between trips, increasing trips per cab by ~2 on average.
• Built a route predictor using graph theory in Python; implemented Dijkstra’s algorithm to allocate minimum idle/dead run routes.
• Designed employee pooling strategy to maximize occupancy per vehicle, minimizing last-mile dead runs and overall fleet cost.
• Delivered measurable cost savings and improved operational efficiency in corporate transport system.

Insurance Policy Claim Likelihood :

• Predicted probability scores for policies indicating their likelihood of getting claimed by the customer.
• Built Logistic regression, Decision tree, Random forest, XGBoost & Simple neural network models in R to predict the scores.
• Had impact of $6 million savings by the Client in one quarter

Artificial Intelligence - Image Processing & Recognition :

• Led a team of 2 data scientists, drafted and implemented the entire Data Modelling part of the project
• Applied various Image Processing techniques such as image gray scaling, hough circle transform, connected components and developed Image Recognition models like CNNs using Opencv, PIL in python
• Got the accuracy of more than 99.5% on Image Recognition models such as SqueezeNet, DarkNet, Normal CNN, etc