My Builds.

FTDUIQA, a fusion model that integrates traditional machine learning and deep learning for assessing underwater image quality. By combining CNNIQA, DBCNN, HyperIQA, MSAEQA, and a modified MAUIQA, our approach significantly improved evaluation accuracy on the UID2021 dataset, outperforming existing methods with a PLCC of 0.8654.

Built an end-to-end real-time big data pipeline that streams Reddit comments through Kafka and a two-stage Apache Flink preprocessing layer for cleaning and validation. Implemented live model retraining with VADER-labeled data and hot-reloading inference, plus an autoscaler that watches Kafka consumer-group lag to dynamically scale preprocessing and inference workers via Docker Compose. Served keyword-driven sentiment scoring and historical backfill through a FastAPI + Redis backend with a Next.js dashboard, fully containerized.

Developed a spatio-temporal traffic forecasting system combining classical baselines (Naive, ARIMA, XGBoost), GRU-based deep learning, and Temporal Fusion Transformer (TFT) for state-of-the-art probabilistic predictions. Implemented MC Dropout and quantile regression for uncertainty quantification and cross-junction dependency modeling. Designed a modular Python codebase with exploratory data analysis and visualization achieving robust RMSE(5.9849) and MAE(4.0017) performance and actionable uncertainty insights for urban traffic planning.

Developed an AI-powered BIM quality assurance platform combining deterministic IFC rule validation and LLM-based analysis using Gemini/OpenAI/MiniMax. Integrated buildingSMART Data Dictionary (bSDD) for standardized property checks, terminology normalization, and multi-language support. Implemented readiness scoring, auto-fix capabilities, and detailed reporting for IFC models, reducing property and naming inconsistencies by ~40%. Designed a modular FastAPI + Next.js architecture enabling scalable validation pipelines, interactive dashboards, and exportable compliance reports.