I am a Machine Learning Engineer focused on building reliable, intelligent systems across healthcare, environmental modeling, and scientific domains. I design robust data pipelines and train/optimize production-grade deep learning models.
- 🎓 Academic Path: M.S. in Data Science at Saint Peter's University, NJ (Expected Feb 2026) | B.Tech in CSE at GITAM University, India.
- 🔬 Specializations: Multimodal Learning, Retrieval-Augmented Generation (RAG), Clinical Informatics, Graph Heuristics, and Explainable AI.
- 💬 Focus Area: Connecting clinical clinical notes with structured EHR data, and refining verification layers in LLMs to enforce factuality.
|
|
🔍 Expand Detailed Badges
Deep Learning & NLP
Computer Vision & Specialized
Languages, Data & Tools
Here is a showcase of my projects, featuring architectural designs and technical summaries. Click to expand each project.
📄 ArgRAG: Argument-Aware Retrieval-Augmented Generation (Click to expand)
Published research introducing ArgRAG, a novel retrieval-augmented generation architecture featuring a validation layer.
- The Problem: LLM generators suffer from hallucination and lack of calibration when reasoning over retrieved knowledge.
- The Solution: Inserts an explicit argumentation classifier between the retriever and the generator to identify the stance of each passage.
- Key Metrics:
- 📈 +3.4% Factual Accuracy on the BEIR-FEVER benchmark.
- 📈 +21.0% Human-Rated Explanation Quality over vanilla RAG.
- 📉 Calibration Error (ECE) decreased to 0.043 (vs. baseline 0.087).
graph TD
Query[User Query] --> Retriever[FAISS Retriever]
Retriever --> Classifier[Argument Stance Classifier <br/> BERT/RoBERTa/DeBERTa]
Classifier --> Map[Map Stances: Supportive / Adversarial / Neutral]
Map --> Generator[LLM Generator <br/> FLAN-T5]
Generator --> Reason[Reason over Arguments]
Reason --> Output[Factual Verdict + Explanation]
style Classifier fill:#4f46e5,stroke:#fff,stroke-width:2px,color:#fff
style Map fill:#8b5cf6,stroke:#fff,stroke-width:1px,color:#fff
style Generator fill:#ec4899,stroke:#fff,stroke-width:2px,color:#fff
- Tech Stack:
BERT,RoBERTa,DeBERTa,FLAN-T5,FAISS,HuggingFace,BEIR
🧬 Multimodal Clinical Outcome Prediction — MIMIC-IV (Click to expand)
An end-to-end pipeline fusing tabular and unstructured EHR data to predict hospital patient outcomes.
- Tabular Branch: Structured clinical database metrics (demographics, lab values) processed via LSTMs and tree-based classifiers.
- Text Branch: Clinical notes and practitioner assessments encoded using BioClinicalBERT embeddings.
- Result: Multimodal fusion improved F1 prediction score by 7% over unimodal baselines.
graph LR
Tabular[EHR Tabular Data] --> TabEncoder[LSTM / RF Encoder]
Notes[Clinical Notes] --> BioBERT[BioClinicalBERT Text Encoder]
TabEncoder --> Fusion[Vector Fusion & Concatenation]
BioBERT --> Fusion
Fusion --> Classification[Output Classifier]
Classification --> Prediction[Outcome Prediction]
style BioBERT fill:#3b82f6,stroke:#fff,color:#fff
style Fusion fill:#8b5cf6,stroke:#fff,color:#fff
- Tech Stack:
PyTorch,Hugging Face,BioClinicalBERT,SQLAlchemy,Scikit-learn
🫁 Chest Radiograph Anomaly Detection (Click to expand)
- Task: Automated detection of chest X-ray abnormalities (pneumonia, tuberculosis).
- Interpretability: Implemented Grad-CAM saliency maps to highlight localized regions influencing prediction scores, providing clinically verifiable rationales.
- Tech Stack:
TensorFlow,Keras,OpenCV,Scikit-learn
🌿 Species Distribution & Environmental Analysis (Click to expand)
- Task: Modeled biodiversity drivers and local habitat profiles.
- Methodology: Applied feature engineering for custom diversity metrics and IQR-based outlier detection to clean sensor noise. Benchmarked Linear Regression, Random Forests, and Gradient Boosting.
- Tech Stack:
Pandas,Scikit-learn,Matplotlib,Seaborn
☄️ Asteroid Diameter Prediction — NASA Dataset (Click to expand)
- Task: Multilayer Perceptron (MLP) regressor predicting asteroid diameters from orbital and physical parameters.
-
Evaluation: Full feature selection and exploratory data analysis evaluated using MAE, MSE, and
$R^2$ metrics. -
Tech Stack:
TensorFlow,Scikit-learn,Pandas
🚚 Delivery Route Optimizer with Real-Time Constraints (Click to expand)
- Task: Real-time delivery routing simulator minimizing delivery time and fuel costs under traffic and window capacity constraints.
- Algorithms: Graph traversal heuristics using Dijkstra, A*, and TSP approximation solvers.
- Tech Stack:
NetworkX,NumPy,Pandas
"Building models is simple; building verification layers that let models self-correct is the real challenge."