Research
Peer-reviewed papers and research systems, spanning real-time LLM serving, continual learning, self-correcting pipelines, and applied deep learning.
SAGE
A learned SLO-aware retrieval policy that picks a per-query passage budget: 95% SLO compliance (vs 30% static), −51% retrieval cost, −36% P95 latency, at a 2-point EM cost.
LLM Blueprint
A blueprint for taking LLMs from prototype to dependable, real-time enterprise deployment.
CAFO
Treat a RAG pipeline as a control system: sense failures, fingerprint them, and apply structural fixes that transfer across queries, delivering 2.6× the correction effect of a per-query oracle.
STAR + FAR
A budget-conscious continual-learning regimen (per-day LoRA adapters with sparse routing plus freshness-aware replay) that gains 3.2–4.1 pp in freshness and legacy retention at ~1.7 min/day update cost.
KAN Tornado
Benchmarking CNN, VGG19, and Xception on the TorNet radar dataset, and showing when replacing MLP heads with KAN layers improves rare-event tornado detection.
NPO Cybersecurity SLR
A PRISMA 2020 systematic review (4,650 records screened) of cybersecurity readiness in non-profits, distilled into actionable, resource-constrained guidance.