Novel AI algorithm design
A state-space model learns healthy update dynamics, isolates suspicious deviations, and repairs them before they can corrupt the global location model.
View CSU STRATA technology pageReal-time graph construction connects a new wireless fingerprint to the most relevant reference locations for accurate positioning without rebuilding or retraining the model.
View CSU STRATA technology pageARMOR uses a dual-AI architecture: one AI handles mobile indoor localization while a second AI watches how the first one learns. When cyberattacks or faulty updates push the locator off course, the guardian detects the error and corrects the learning process before the shared model is damaged.
Read the paperGATE strengthens graph-based localization with two hyperspace representations that improve message passing and reduce the “blind spots” that can hide useful signal relationships from a graph neural network. Its real-time edge-construction policy then rebuilds the graph during inference, letting the model adapt as Wi-Fi conditions change.
Read the paperLogNet performs localization with a network built entirely from binary logic gates—AND, OR, NAND, NOR, and XOR—operating on binarized Wi-Fi signal strengths. Its traceable gate decisions show which access points drive each reference-point prediction and reveal how long-term signal changes cause errors, while keeping the model compact and fast.
Read the paperDAILOC learns from new phones and changing environments without retraining from scratch or forgetting earlier conditions. A multi-level variational autoencoder separates device and temporal shifts from location-relevant features, while memory-guided class-latent alignment preserves previously learned locations as new domains arrive.
Read the paperCIELO lets an indoor-localization model add new reference points over time without retraining from scratch or losing accuracy on places it already knows. It expands the location label space through representation-memory replay and selective sampling, retaining a compact set of useful past fingerprints to limit catastrophic forgetting.
Read the paperSAFELOC is a federated indoor-localization framework designed specifically to withstand poisoned training data. One compact network detects poisoned fingerprints, cleans them, and estimates location, while saliency-based aggregation reduces each client model’s influence according to the severity of the detected attack.
Read the paperSENTINEL modifies capsule neural networks so indoor localization remains accurate when adversarial or rogue “evil-twin” access points manipulate Wi-Fi fingerprints. It also introduces RSSRogueLoc, a real-world rogue-access-point dataset used to test defenses beyond simulated attacks.
Read the paperCALLOC trains against adversarial Wi-Fi disruptions through an adaptive curriculum, presenting easier attacks first and increasing their difficulty as the model improves. A lightweight scaled dot-product attention network focuses on the most useful signal relationships while keeping the defense practical for mobile devices.
Read the paperSTELLAR combines contrastive learning with a Siamese multi-headed attention network to learn which Wi-Fi fingerprints belong to the same place despite device differences and signal drift over time. By tackling both sources of variation together, it transfers to new phones and later measurements without recalibration or retraining.
Read the paperFedHIL lets phones train a shared indoor-localization model without sending their raw Wi-Fi fingerprints to a server. Its selective weight-adjustment method aggregates only the relevant model tensors, limiting the damage from extremely noisy devices while preserving useful knowledge from other clients.
Read the paperVITAL is the first indoor-localization framework to turn Wi-Fi fingerprints into image-like representations and process them with a vision transformer, emphasizing access points that remain visible across devices. Its reusable data-augmentation module helps the model generalize to unseen phones without device-specific calibration.
Read the paperSANGRIA uses a greedily trained stacked autoencoder to generate synthetic Wi-Fi fingerprints that imitate how signal readings vary across devices. It combines those augmented fingerprints with real measurements and trains a categorical gradient-boosting model, improving cross-device localization with modest computing cost.
Read the paperThe ANVIL framework combines a lightweight multi-head attention network with data augmentation designed specifically for device-invariant Wi-Fi fingerprinting. Its attention heads learn complementary signal relationships, enabling calibration-free localization across phones with different radio hardware.
Read the paperThis work combines ground-penetrating-radar signal processing with image analysis to turn subsurface echoes into measurable features. It isolates the characteristic reflection produced by a buried object and uses its geometry to approximate the object’s size without excavation.
Read the paperSenior Silicon Design Engineer at AMD · Austin, Texas
Computer Scientist
AI researcher and silicon performance optimization engineer working across AI algorithm design, AI model security, explainable AI, computer architecture, performance optimization, edge intelligence, and reliable indoor positioning. My work turns complex systems into models that can be understood, optimized, and deployed.
Research & industry affiliations
Forthcoming from Springer Nature.
ARMOR first learns what a normal, healthy update to its central location model should look like by using a special “state-space model” (SSM). This SSM acts like a smart predictor, anticipating the next expected changes in the model’s “brain.” When a mobile device sends its updated location data, ARMOR checks if these updates match what the SSM predicted. If the updates deviate too much from the prediction, ARMOR steps in. It figures out exactly which parts of the update are problematic and adjusts them, essentially cleaning up the data before it can corrupt the main location model. This technique means only reliable and corrected information is used to improve the overall location system.
View CSU STRATA technology pageGATE (Graph Attention Neural Networks with Real-Time Edge Construction) represents indoor environments as a graph of interconnected reference locations, learning how neighboring locations influence one another instead of processing each wireless fingerprint in isolation. During inference it dynamically connects each newly observed fingerprint to its most relevant region of the reference graph — no rebuilding or retraining — enabling accurate, real-time indoor localization on commodity smartphones with low computational overhead.
View CSU STRATA technology pageAI innovation and technical leadership in silicon performance engineering.
Scientific Research Honor Society.
Explainable AI for indoor localization using logic-gate interpretation.
Graph attention networks for robust indoor positioning.
Company recognition for AI algorithm development.
Attention-based learning for device-invariant indoor localization.
AI-driven pre-silicon performance modeling and optimization for complex memory-system architectures.
Performance analysis, tooling, and design-space exploration across memory controllers and DRAM subsystems.
Research role designing secure, compact, and explainable AI systems for indoor localization on edge devices.
Built software for remote interaction with FPGA accelerator hardware and workload execution.
Designed, configured, and secured enterprise network infrastructure and systems.
Developed backend functionality and reliable device-to-server communication for embedded systems.
Designed and tested embedded electronics, including PCB and circuit integration.
Research in AI algorithms for robust indoor localization and resource-constrained edge systems.
Research in AI algorithms for robust indoor localization and resource-constrained edge systems.
Undergraduate research in signal processing using ground-penetrating radar (GPR).
For research collaboration and technical conversations across AI systems, silicon performance, edge intelligence, and reliable localization.
Google Scholar
LinkedIn
ORCID
Web of Science
IEEE Xplore