I build and evaluate LLM-based systems: retrieval-augmented generation, parameter-efficient fine-tuning, multi-agent pipelines, and the monitoring/uncertainty infrastructure that keeps them trustworthy in production. My work leans empirical: most repositories below report benchmark numbers rather than just claims, because I care about results that replicate.

I'm currently preparing applications for research-oriented graduate study (MS/PhD), with interests spanning uncertainty quantification, model evaluation methodology, and efficient adaptation of large models. Open to research collaborations, co-authored write-ups, and conversations with labs working in these areas.

• Uncertainty & confidence estimation in deep networks (label-free and unsupervised methods)
• Evaluation methodology for LLM/RAG systems (faithfulness, hallucination, retrieval quality)
• Parameter-efficient fine-tuning and adaptation of large models
• Multi-agent LLM systems with verification / critique loops
• Production ML monitoring: distribution drift, model degradation over time

Evaluation & Uncertainty

• rag-evaluation-framework: Faithfulness, hallucination, retrieval precision, and answer relevance metrics for RAG pipelines, including UCM, an unsupervised confidence metric that requires no ground-truth labels.
• unsupervised-confidence-estimation: Benchmarks MSP, MC Dropout, Evidential Deep Learning, and Deep Ensembles against a novel label-free uncertainty metric on CIFAR-10.
• production-drift-detection: KL divergence, PSI, MMD, and ADWIN drift detectors with empirical benchmarks and a FastAPI monitoring dashboard.

Model Adaptation

• parameter-efficient-fine-tuning: QLoRA fine-tuning across 2B-70B parameter LLMs; on small-data regimes, LoRA reaches 21% better perplexity than full fine-tuning while updating 421x fewer parameters (full benchmark setup and baselines in repo).
• multi-objective-feature-selection: NSGA-II based feature selection on medical tabular data; 9 of 30 features match full-feature baseline accuracy (94.74%) at 70% dimensionality reduction.
• loss-landscape-analysis: Empirical comparison of MSE vs. cross-entropy on MNIST, with gradient norm analysis explaining the convergence gap.

Multimodal & Applied Systems

• multimodal-medical-vqa: Cross-attention fusion of BioViL-T and Mistral-7B (QLoRA) for clinical visual question answering, with MC Dropout confidence and Grad-CAM explainability.
• multi-agent-research-system: Researcher/Analyst/Critic agent loop in LangGraph with a hallucination and citation validation gate; cached-search retrieval measured at a 10,184x speedup over uncached calls in repeated-query benchmarks.