Publications & Pre-prints

HQ-Mamba-PAC is a hybrid quantum-classical vision framework that combines semantic distillation, quantum encoding, messenger-linked quantum processing, selective state-space fusion, and PAC-Bayesian regularization in one pipeline. The goal is to keep the long-range modeling benefits of advanced vision systems while improving trainability, robustness under domain shift, and theoretical control over generalization, including deployment for non-invasive chick gender classification.

Omni-PAIFT is a unified multimodal foundation architecture designed to reduce information dilution, attention sinks, gradient interference, and expensive uncertainty estimation in large-scale systems. It combines threshold-gated mixture-of-depths differential attention, cross-modal fusion, single-pass uncertainty routing, evolutionary multi-task optimization, and speculative sparse execution to improve long-context efficiency and downstream task performance with only modest additional compute.

This paper presents a real-time gesture-to-music generation system that turns continuous human motion into expressive audio rather than mapping gestures to fixed pre-recorded outputs. It uses Liquid Time-Constant networks for low-latency motion understanding, an audio-visual mixture-of-experts router for latent control, and flow-matching or diffusion-based generation to synthesize music and vocals on the fly while aligning the system with human aesthetic intent through GDPO.

CareerCompass is a privacy-first employability prediction framework that allows multiple institutions to train a shared model without centralizing sensitive student records. It combines federated learning for decentralized prediction, propensity-score-based causal inference to test whether suggested interventions truly improve employability, and GraphRAG over a knowledge graph to generate personalized learning paths tailored to each student's current skills and career goals.

This work proposes a federated brain tumor segmentation framework that turns model failure patterns into anonymous causal error vectors instead of sharing raw medical images. A server-side latent diffusion model then generates synthetic hard examples that specifically target those blind spots, while MO-DGPO and TreeGRPO make the optimization more stable and reduce generation overhead for practical federated training.

Cerebro-TabM is a clinical stroke prediction architecture built for highly imbalanced and partially missing healthcare data, where rare positive cases and missing-not-at-random features can mislead standard tabular models. It introduces a missingness-aware tokenizer, an evidential BatchEnsemble that outputs uncertainty-aware Dirichlet evidence, and a PAC-Bayes-grounded focal loss that aims to improve calibration, fairness, and edge-deployment readiness for safety-critical decision support.

This architecture is designed for tabular learning with mixed numerical and categorical features, using a shared embedding space followed by cross-attention to model interactions between feature types. It further applies squeeze-and-excitation recalibration, lightweight feedforward blocks, and efficiency techniques such as blockwise attention, sparse masking, pruning, and quantization to improve performance without making the model too expensive to train or deploy.

HelixVantage is a hierarchical medical AI orchestration system that grounds language-based reasoning in live physiological signals rather than relying only on static text retrieval. It converts edge sensor streams into HL7 FHIR observations, combines those signals with user queries through a context-aware router, and applies a safety-oriented regularizer so the final response stays closer to clinical guidelines while reducing hallucinations.

BrepHCC is an unsupervised 3D CAD clustering framework built for large jewelry design repositories, where visual similarity alone is not enough to capture style, topology, and manufacturing structure. It fuses native Rhino B-Rep topology with dense point-cloud features through a lightweight top-2 mixture-of-experts router and trains with a hierarchical contrastive clustering loss that preserves geometric detail, curvature patterns, and cross-model similarity.

PLICI is a deterministic security middleware for LLM agents that shifts defense away from fragile prompt-level heuristics toward verifiable protocol enforcement. It uses session-scoped Ed25519 signatures to verify data origin, schema-locked enclaves to isolate tool payloads before the agent reads them, and an NLI-based semantic filter to catch logic hijacking, sharply reducing attack success while keeping latency low enough for real-time agent systems.