Research agenda

Efficient systems. Robust evidence. Reproducible conclusions.

My research asks not only whether a machine-learning method performs well, but whether its gains remain reliable under controlled evaluation, distributional perturbations, computational constraints, and realistic deployment conditions.

principles.py rigor + efficiency + robustness + transparency

Research map 01 Efficient NLP 02 Trustworthy Evaluation 03 Higher-Order Learning 04 Explainable & Applied AI

Efficient NLP & LLM adaptation

Reducing model cost without disguising the quality trade-off

I study knowledge distillation, low-rank adaptation, quantization, gradient checkpointing, adaptive rank control, and compression-aware training to determine how much model capacity and computation are actually necessary for reliable NLP systems.

Current questions

How much model capacity is necessary for a target task?
Which parameter-efficient fine-tuning components produce measurable gains under matched computational budgets?
How should predictive quality, VRAM, runtime, convergence, and optimization stability be reported together?
Can compact models preserve useful reasoning and evidence, rather than only prediction outputs?

Current work

Ultra-efficient fine-tuning of Qwen2.5-7B-Instruct on a single consumer GPU using 4-bit loading, QLoRA, adaptive rank scheduling, SVD warm-start, gradient-noise-gated control, and reproducible GPU isolation.

Knowledge Distillation QLoRA 4-bit Quantization Adaptive Rank Memory Profiling Efficient LLMs

Trustworthy evaluation & optimizer reliability

Measuring whether optimization gains survive beyond clean accuracy

I treat experimental design as part of the research contribution. My work investigates whether apparent performance improvements remain credible after controlling hyperparameters, data access, leakage, random variation, corruption, label noise, calibration, computational cost, and evaluation methodology.

Evaluation principles

Separate model selection from final held-out evaluation.
Match data access, tuning opportunities, optimization budgets, and computational resources.
Stress-test models under feature corruption and label noise.
Quantify sensitivity to hyperparameters, seeds, and optimization schedules.
Report calibration, runtime, convergence, and computational cost alongside predictive accuracy.

Final-stage research

Beyond Clean Accuracy: A Test-Blind Benchmark for Optimizer Reliability, Robustness, Calibration, and Computational Efficiency

I am developing a test-blind optimizer benchmark across 12 tabular classification datasets that evaluates neural network optimization algorithms beyond conventional clean test accuracy. The framework measures optimization stability, robustness to data corruption and label noise, hyperparameter sensitivity, predictive calibration, loss-landscape geometry, and computational efficiency.

Optimizer Reliability Test-Blind Evaluation Robustness Calibration Hyperparameter Sensitivity Loss Landscapes Label Noise Ablation Studies Statistical Testing Reproducibility

Graph & hypergraph learning

Finding when higher-order structure is genuinely useful

I compare pairwise graph projections with incidence-based hypergraph representations to determine when multi-entity interaction structure provides measurable improvements in biological structure recovery.

Benchmark design

Nine IntAct MITAB interaction datasets.
Held-out, leakage-controlled evaluation.
Matched-k clustering comparisons.
Penalized hypergraph spectral clustering.
Null models, significance testing, and Gene Ontology enrichment.

Research objective

Move beyond the broad claim that higher-order representations are inherently superior and instead characterize the empirical conditions under which incidence structure contributes measurable information beyond pairwise projection.

Hypergraphs Spectral Clustering Protein Complexes Network Science Null Models GO Enrichment Leakage Control

Explainable AI, software security & AI for science

Building efficient models that expose the evidence behind their decisions

My applied work extends efficient and trustworthy machine learning to software security, computational biology, digital phenotyping, healthcare AI, and sustainable computing. A central direction is designing compact systems whose predictions remain interpretable without depending on expensive post-hoc explanation pipelines.

Final-stage research

ARGUS-KD: Attribution-Rationale Grounded Unified Student Distillation for Explainable Code Vulnerability Detection

ARGUS-KD investigates explainable knowledge distillation for code vulnerability detection by transferring not only a teacher model's prediction outputs but also its underlying rationale and security evidence. Commit patch diffs are used to derive ground-truth vulnerability rationales for training a compact student model.

Explainability by design

Distill predictive behavior together with security rationales.
Use commit patch diffs as supervision for vulnerability evidence.
Train a dedicated rationale prediction head.
Introduce a security concept bottleneck for structured evidence.
Produce token-level vulnerability explanations during inference without additional explanation-time computation.

Broader application areas

Explainable code vulnerability detection.
Computational biology and biological network analysis.
Longitudinal and naturalistic digital phenotyping.
Sustainable AI across energy, carbon, water, and hardware lifecycle.
Healthcare NLP under uncertainty and deployment constraints.

Long-term direction

Developing machine-learning systems that are efficient enough to deploy, robust enough to survive realistic variation, interpretable enough to audit, and experimentally rigorous enough to trust.

ARGUS-KD Explainable AI Knowledge Distillation Code Vulnerability Detection Token-Level Rationales Concept Bottlenecks AI4SE Computational Biology Digital Phenotyping Sustainable AI

How I work

A research workflow built for credible comparison

Specify the claim

Define the exact quality, efficiency, robustness, interpretability, or representation question before optimizing the system.

Control the comparison

Match budgets, data access, preprocessing, evaluation splits, hyperparameter opportunities, and computational resources.

Stress-test the result

Use repeated runs, corruption tests, label noise, ablations, null models, sensitivity analysis, calibration measures, and statistical validation.

Make it reproducible

Pin environments, log configurations, isolate hardware, preserve experimental artifacts, report failure modes, and separate model-selection decisions from final evaluation.