All Projects

Complete portfolio of AI/ML projects demonstrating expertise across various domains.

Consulting Projects

Industry projects delivering real-world business impact

GenAI for Asset Integrity and Process Safety

Completed

Mitigated risk of $10M+/year due to safety incidents at ADNOC by developing a comprehensive RAG agent with automated graph generation and LLM-based analysis for hazardous events.

RAG Text2SQL LLMs BERT Azure ML Docker
$10M+/year risk mitigation

Key Achievements:

  • Developed RAG agent with automated graph generation on top of Text2SQL engine to visualize latest trends
  • Utilized LLMs for trend analysis and key insights on hazardous events using prompt engineering & data manipulation
  • Devised control KPIs to indicate likelihood of hazardous events at each equipment and location in various assets
  • Performed benchmarking of pre-trained LLMs and fine-tuned BERT models for text-classification on Oil&Gas data
  • Deployed dockerized applications on AzureML and presented demo at ADIPEC 2024

Recommendation System for Retail

Completed

Developed recommendation engine for Jack in the Box (USA) to suggest top 3 relevant products to consumers, integrating customer profiling with collaborative filtering techniques.

Collaborative Filtering Matrix Factorization K-Means Clustering DBSCAN Data Warehouse

Key Achievements:

  • Reconciled datasets from customer engagement, loyalty and marketing tools to build comprehensive data warehouse
  • Evaluated various recommendation techniques based on collaborative filtering and matrix factorization
  • Profiled customers using KMeans Clustering and DBSCAN algorithms
  • Integrated customer profiles with recommendation model for personalized product suggestions

Vision Models for Real-Time Inspection

Completed

Minimized safety risks and operation downtime for PTFI through real-time anomaly detection models deployed on edge devices with optimized 4 fps latency using TensorRT.

Computer Vision GLCM Gabor Filters Autoencoder Siamese Network TensorRT Edge Computing
4 fps real-time inference

Key Achievements:

  • Explored efficacy of features using GLCM, Gabor filters, power spectral density and optical flow algorithms on mineral ore images
  • Analyzed texture, volume and flow rate properties of mineral ore through advanced image processing
  • Trained and fine-tuned anomaly detection models using Autoencoder and Siamese network architectures
  • Deployed final model on GPU-powered edge computing devices using TensorRT for optimized latency of 4 fps

Physical Modelling for Product Quality Forecast

Completed

Increased molybdenum mineral production by 3% through Python-based digital-twin modeling of chemical plants with automated optimization and ML-based quality forecasting.

Digital Twin Python L-BFGS-B Random Forest RNN CNN
3% production increase

Key Achievements:

  • Developed Python-based digital-twin model of three chemical plants to predict impurity concentration in the product
  • Automated digital-twin performance optimization using methods like L-BFGS-B, Nelder-Mead and trust-region
  • Analyzed ML techniques like Random Forest, Boosted Trees, RNNs and CNNs to forecast product quality
  • Achieved 3% increase in molybdenum mineral production through optimized process control

Mining Operation Optimization

Completed

Increased production hours of mining operations by 8% through Kalman Filter-based mineral concentration estimation and CNN-powered obstruction detection with 98% accuracy.

Kalman Filter CNN Computer Vision Azure ML Edge Devices
8% production hours increase

Key Achievements:

  • Implemented and fine-tuned Kalman Filter to accurately estimate mineral concentrations in absence of sensors
  • Trained CNN-based image classification model to detect obstruction in mineral ore crushers with 98% accuracy
  • Deployed solutions across edge devices for real-time processing
  • Integrated AzureML toolkit for real-time monitoring and performance tracking

Customer Churn Analysis for Teleservices

Completed

Analyzed over 300 GB of network data to identify factors contributing to customer churn at NSight Teleservices, utilizing ML classification models to predict churn likelihood.

Big Data Random Forest XGBoost Network Analysis Churn Prediction
300 GB data analyzed

Key Achievements:

  • Analyzed over 300 GB of network performance and traffic data to identify hotspot towers serving churned customers
  • Conducted competitor analysis on offered services and correlated inaccuracies in billing with poor customer service
  • Explored efficacy of classification models like Random Forest and XGBoost to represent churn likelihood of customers
  • Delivered insights to improve customer retention strategies and service quality

Research Projects

Academic research and cutting-edge AI exploration

Deepfake Detection and Inference Optimization

Completed

Optimizing deepfake detection models for deployment on H100 GPUs at A*STAR's Institute for Infocomm Research (I2R), focusing on model quantization, custom CUDA kernels, and benchmarking performance across diffusion-based datasets.

PyTorch CUDA Model Quantization H100 GPUs Vision Transformers AWS SageMaker

Key Responsibilities:

  • Optimizing deepfake detection models for deployment on H100 GPUs by profiling call-stacks to reduce CPU overhead
  • Reducing model size through experiments with various quantization techniques
  • Benchmarking model performance using various diffusion-based datasets to track performance loss during optimization
  • Implementing fast-attention with custom CUDA kernels to accelerate attention calculation for vision transformer models

Organization: Institute for Infocomm Research (I2R), A*STAR, Singapore

Status: Completed

Contextual Bandit-Over-Bandit Optimization

In Progress

Research on whether an online algorithm for policy selection can adapt to unknown reward function shapes without prior knowledge of which policy learns the best. A meta-bandit/expert observes MSE (or Doubly Robust MSE) signal to adaptively select from a portfolio of arm selecting policies — linear (LinUCB/LinTS), polynomial, and NeuralTS (NTK-at-init features) — each round, to consistently outperform any single fixed policy across diverse reward functions.

Contextual Bandits Online Meta-Learning Gaussian Process Neural Tangent Kernel Thompson Sampling scikit-learn MLflow PyTorch

Key Responsibilities:

  • Designed a bandit-over-bandit framework where a meta-bandit (Hedge, AdaHedge, GP-UCB, GP-TS) adaptively selects among a portfolio of inner learner configs each round
  • Built modular estimators — LinUCB/LinTS (Sherman-Morrison V⁻¹), polynomial expansion, GP (RBF + Woodbury inverse), and NeuralTS (NTK-at-init features) — with a unified interface
  • Supported bandit policies (Thompson Sampling, UCB, Greedy) and benchmarked across reward function types (linear, square, cubic, sine, cosine, MLP)
  • Evaluated online meta-learner against all fixed offline configs across multiple seeds, tracking cumulative regret with mean ± SEM plots
  • Integrated MLflow with per-round CSV traces, debug arm logs, and artifact versioning for fully reproducible experiment tracking

Organization: M3S, Singapore-MIT Alliance for Research and Technology (SMART)

Status: Ongoing

Quantum Computing Exploration

Completed

Explored quantum computing fundamentals by executing basic QC programs and comparing performance across different quantum hardware platforms on AWS and Azure cloud services.

Qiskit Q# Braket Pennylane AWS Azure

Key Achievements:

  • Executed basic quantum computing programs such as Bell state and classification using different libraries
  • Implemented quantum algorithms using Qiskit, Q#, Braket, and Pennylane frameworks
  • Compared cost and performance of various quantum hardware by Rigetti and IONQ
  • Evaluated quantum computing services on AWS and Azure cloud platforms

Organization: WWT R&D Team

Read Article

Federated Learning Methodologies

Completed

Conducted comprehensive analysis of federated learning approaches including Auto-Fed averaging and Hyper-Networks with personal classifier for image classification on CIFAR-10 dataset.

Federated Learning Auto-Fed Hyper-Networks CIFAR-10 Image Classification

Key Achievements:

  • Conducted analysis of two federated learning methodologies: Auto-Fed averaging and Hyper-Networks with personal classifier
  • Implemented and evaluated both approaches for image classification on CIFAR-10 dataset
  • Analyzed methodologies on KPIs indicating generalization, personalization and robustness
  • Compared performance metrics across different federated learning strategies

Organization: WWT R&D Team

AIOps Framework for IT Operations

Completed

Developed modular and dockerized ML framework for quick deployment of ML applications such as failure detection and load forecasting in IT servers at scale, integrated with popular IT databases.

Python ADTK ARIMA Docker Splunk Prometheus ETL

Key Achievements:

  • Developed modular and dockerized ML framework using Python libraries ADTK and ARIMA
  • Enabled quick deployment of ML applications for failure detection and load forecasting in IT servers at scale
  • Integrated the solution with ETL toolkit to query data from popular IT databases like Splunk and Prometheus
  • Created scalable infrastructure for ML-driven IT operations monitoring

Organization: WWT R&D Team

Read Article

Data Center Power Optimization

Completed

Trained neural network model to estimate power utilization effectiveness (PUE) of data centers and conducted optimization analysis using SHAP values to reflect impact of cooling parameters.

Neural Networks Optuna SHAP Power Optimization Data Center

Key Achievements:

  • Trained neural network model to estimate power utilization effectiveness (PUE) of data centers as a function of cooling system parameters
  • Optimized model architecture using Optuna for hyperparameter tuning
  • Conducted PUE optimization analysis on SHAP values of trained model reflecting impact of cooling parameters
  • Delivered insights for improving energy efficiency in data center operations

Organization: WWT R&D Team

Read Article

Speech and Video Refinement

Completed

Developed ML solution to detect vocal disfluencies in speech using MFCC features and regenerate refined video frames through optical flow-based interpolation, achieving 90% accuracy.

Random Forest MFCC Optical Flow Audio Processing Video Processing
90% accuracy

Key Achievements:

  • Trained Random Forest classification model to tag vocal disfluencies in speech using MFCC features
  • Achieved 90% accuracy in detecting vocal disfluencies from audio data
  • Regenerated frames with vocal disfluencies using optical flow-based frame interpolation
  • Produced refined video output with seamless audio-visual synchronization

Organization: WWT R&D Team

Academic/Course Projects

University coursework and learning projects

Mobility Assistant for Visually Impaired (MAVI)

Completed

Developed an assistive device at IIT Delhi under Prof. M. Balakrishnan to help visually impaired individuals navigate their environment using real-time object detection with audio feedback.

YOLO-v2 Raspberry Pi Python Text-to-Speech Movidius Neural Stick
8 fps real-time, 70% CPU reduction

Key Achievements:

  • Deployed pretrained YOLO-v2 object detection model on Raspberry Pi chip for real-time detection
  • Integrated text-to-speech capabilities with object detection output in Python for audio feedback
  • Optimized latency to 8 fps by integrating Movidius neural stick to reduce CPU load by 70%
  • Created an accessible mobility solution for visually impaired individuals

Timeline: Jan 2019 - April 2019

Supervisor: Prof. M. Balakrishnan, IIT Delhi