Thathsara Rajapaksha

End-to-end MLOps pipeline on Azure: ingests financial data, trains LightGBM models, and serves real-time predictions via serverless API. Cost-optimized architecture scales from $0 to production workloads.

Azure ML Python Blob Storage SHAP

Engineered an enterprise-grade credit risk system combining LightGBM predictive modeling with Llama 3.2 for automated loan decisions. Features big data streaming with Polars, actuarial scorecard calibration, and real-time AI explanations—processing 2.2M+ records with <1GB RAM.

Python LightGBM Polars Llama 3.2 Streamlit Scorecardpy

End-to-end fraud detection system featuring automated experiment tracking, model registry management, and production-grade drift monitoring.

Python MLflow Docker Flask

Developed a Lambda Architecture-based system for analyzing IoT energy consumption data with batch and real-time processing for anomaly detection.

Apache Hadoop Apache Spark Kafka Docker

Analyzed 27,500 booking records to identify patterns in cancellations and no-shows, optimizing revenue and resource planning.

Python Scikit-learn Data Analysis Customer Segmentation

A Neuro-Symbolic financial agent that orchestrates deterministic tools (Prophet, Risk Metrics) with LLM reasoning. Features autonomous multi-asset analysis, ensemble forecasting, and hybrid execution modes.

Python LangGraph Streamlit Docker Gemini

An autonomous AI agent that writes, debugs, and executes Python code to analyze raw data. Features a cyclic "Reason-Act-Observe" architecture using LangGraph and Gemini 2.5.

Python LangGraph FastAPI Docker Gemini

A complete AutoML system training 6 models (XGBoost, LightGBM, CatBoost) with Optuna optimization, achieving 86.89% accuracy. Split architecture: Colab training + Streamlit deployment.

Python Colab XGBoost Streamlit

Implemented causal inference T-Learner to optimize email campaigns, achieving 123% conversion improvement while reducing costs by 70% through intelligent customer targeting.

Python XGBoost Causal Inference Uplift Modeling

A hybrid AI system for precision marketing. Combines Unsupervised Learning (K-Means) for segmentation with Supervised Learning (XGBoost) to predict campaign buy-in with 90% AUC.

Python Scikit-Learn XGBoost Looker

High-performance financial sentiment analysis using a Quantized (Int8) DistilBERT model. Reduced model size by 75% and boosted CPU inference speed by 3.5x using ONNX Runtime.

Python Hugging Face ONNX Gradio

An end-to-end recommendation engine featuring a decoupled microservices architecture. Includes XGBoost inference, SHAP explainability, and automated CI/CD pipelines.

Python FastAPI CI/CD

A progressive collection of NLP projects mastering LLM adaptation, from full BERT fine-tuning to parameter-efficient LoRA (95% reduction).

Python PyTorch Hugging Face

A comparative study of traditional ML and Transformer architectures (DistilBERT) on the SQuAD dataset, achieving 98% accuracy in question classification.

Python Hugging Face Scikit-learn

Developed an interactive sentiment analysis application that combines RoBERTa and VADER technologies with real-time visualizations.

Python Hugging Face Flask

Conducted comprehensive exploratory data analysis on Colombo's rental apartment market, analyzing 250+ property listings.

Python Pandas Matplotlib Scikit-learn

Utilizing transfer learning with MobileNetV2/VGG16 architectures to identify brain tumors in MRI scans.

Python TensorFlow Transfer Learning Computer Vision

Developed and deployed a Streamlit-based web application that uses LSTM models to predict stock prices for multiple brands.

Python Streamlit LSTM

User-friendly web application using Streamlit and Google's Generative AI to enhance resume quality across different experience levels.

Python Streamlit Generative AI NLP

Analyzing user reviews from Amazon to determine sentiment towards products using pre-trained BERT model.

Python BERT NLP Hugging Face

Predicts customer churn using XGBoost, Random Forest, and Logistic Regression with hyperparameter tuning.

Translated business requirements into interactive dashboards for real-time tracking of key performance indicators (KPIs).

A model that uses KMeans algorithm for customer segmentation by implementing the Elbow method.

Python Scikit-learn KMeans Clustering