A curated showcase of data-driven and AI-powered projects spanning research, engineering, and agentic intelligence.
Built a semi-autonomous productivity assistant using LangChain and OpenAI APIs for task management and focus cycles. Demonstrated how lightweight agentic design can support context persistence and personalized task flow in everyday productivity environments.
Developed a reasoning agent that analyzes business briefs, financial data, and reports using vector search and LLM-driven synthesis. Streamlined multi-document understanding for early-stage consulting scenarios and automated structured report generation.
Created a semantic Q&A app using FAISS and Hugging Face Transformers, enabling retrieval with page-level grounding. Enhanced explainability and trust in document-based LLM outputs while showcasing reproducible RAG pipelines.
Designed a Dialogflow CX assistant for healthcare appointments and triage with webhook integration for dynamic scheduling. Reduced manual scheduling steps and demonstrated scalable conversational AI design for domain-specific workflows.
Built an end-to-end anomaly detection pipeline using Autoencoder and Isolation Forest models with CI/CD orchestration in Azure ML. Automated log triage and established a reusable template for integrating ML models into enterprise audit pipelines.
Processed transactional datasets using PySpark and built Power BI dashboards for sales, margin, and fraud insights. Improved data accessibility and enabled leadership to track business metrics in near real-time.
Conducted end-to-end analysis of user engagement patterns on digital platforms. Applied logistic regression model to reveal purchase intent drivers and provided optimized campaign recommendations for effective user engagement.
Modeled ICU mortality using Random Forest and SHAP-based feature attribution on clinical variables. Highlighted interpretable predictors for sepsis outcomes and demonstrated the value of explainable ML in clinical settings.
Clustered >5,000 ICU patients to uncover clinically distinct hypotension subgroups with outcome variations. Improved understanding of patient heterogeneity, informing future data-driven risk stratification research.
Fine-tuned ViT-B16 for binary classification using PyTorch AMP and modular training loops. Provided a reproducible baseline for transformer-based computer-vision models on limited datasets.
Implemented and compared Multi-Armed Bandit strategies (ε-greedy, UCB, Thompson Sampling). Demonstrated adaptive policy behavior under uncertainty, bridging classical RL with decision-support systems.
Automated preprocessing and annotation parsing for mosquito larvae images using LabelMe JSON pipeline. Standardized dataset preparation, improving downstream model consistency for vector-borne disease studies.