CRISPR-based diagnostics are emerging molecular tools being explored to improve rapid tuberculosis (TB) detection in low-resource settings. However, adoption in field settings is limited by the need for a cold supply chain and trained personnel. Our work develops a CRISPR-Cas13a–responsive DNA/RNA hydrogel as a robust, sensitive bead-based platform for point-of-care TB detection. In this system, Mtb-specific crRNA complexes with Mtb genomic RNA, activating Cas13a trans-cleavage of Poly-rU motifs. Following activation, Cas13a degrades Poly-rU RNA crosslinks in alginate beads, driving a visible hydrogel-to-liquid transition. We link a reporter molecule to these DNA/RNA linkers to generate a rapid, visible colorimetric output. We will optimize monomer and DNA crosslink concentrations, storage temperature, and buffer conditions to maximize detection robustness while preserving Cas13a activity, and evaluate lyophilization effects on long-term bead stability and functional recovery. Additionally, we will assess the sensitivity, specificity, and limit of detection of our Mycobacterium tuberculosis diagnostic assay in vitro, followed by performance testing in clinical samples. These parameters will be benchmarked against current gold standards such as GeneXpert. By leveraging the stability and novelty of DNA/RNA alginate gels, we aim to deliver a highly sensitive CRISPR-Cas13a assay that provides an economical, point-of-care friendly TB test and expands access to rapid, accurate diagnosis in decentralized, resource-limited settings.

Desalination and water purification via reverse osmosis (RO) and nanofiltration (NF) membrane technologies is a critical technology for producing potable water quality from brackish water, and for reuse applications via treatment of industrial and municipal wastewater, as well as agricultural drainage water. In particular, the combination of RO and NF membrane elements can provide for flexible RO/NF (FLERONF) operation to reduce the required feed pressure. Moreover FLERO operation in the model of partial concentrate recycle should allow desalination over wide-ranges of both product water recovery and feed water salinity. Accordingly, the present project focus is on evaluation the potential for reduction in energy consumption for a range of water salinity and productivity. Accordingly, both the steady-state and transient modes of FLERO operation were modeled, via data-driven machine learning (ML) models, based on extensive experimental FLERO data. The utility of the ML models for prediction of permeate flux and salinity will be presented, along with exploration the path for energy-optimal operation.

This project develops a dashboard to visualize global Starlink outage patterns over time using both historical and real-time network measurement data. The central research question is how outage activity in low-Earth orbit (LEO) satellite networks is distributed geographically, how it evolves across user-defined time windows, and what factors contribute to observed disruptions. The system integrates datasets ranging from CSV-based outage logs to data stored in the lab server, computing normalized outage ratios at country and city-level granularity. The architecture combines a FastAPI backend for data aggregation and API services with a React and Leaflet frontend that renders choropleth maps using ISO country codes. Methodologically, the project applies principles from distributed systems, data modeling, and networking to design an efficient pipeline for querying and processing high-volume, time-series data. Results show that outage activity can be effectively summarized through normalized metrics, revealing spatial disparities and temporal trends in network reliability. As a minimum viable product MVP, the system establishes a functional end-to-end pipeline for data ingestion, processing, and visualization, while providing a foundation for iterative development. These findings support further investigation into correlations with external factors such as geographic coverage and environmental conditions. Future work will extend the platform by incorporating more historical data, improving real-time analytics capabilities. The

High-Level Synthesis (HLS) enhances productivity by generating hardware from C/C++, but optimal performance remains highly dependent on the manual, combinatorial challenge of tuning synthesis directives. While automated Design Space Exploration (DSE) frameworks like AutoDSE address this, they are often architecturally coupled to FPGA-oriented toolchains, of which Xilinx Vitis is an example. The work this quarter was to retarget AutoDSE to the Siemens Catapult HLS environment, an ASIC-oriented tool with distinct project structures, pragma syntaxes, and reporting mechanisms. We developed a scripted TCL-based execution path to automate Catapult’s GUI-centric workflow and introduced a translation layer to map AutoDSE’s directive markers to Catapult-specific pragmas and auxiliary Tcl directives. Furthermore, the framework’s analysis layer was redesigned to extract latency and bottleneck data from Catapult’s cycle.rpt and loop execution profiles. Experimental validation on nested-loop kernels, such as a 2mm matrix multiplication, demonstrated latency reductions of up to 50% within a ten-iteration budget. Although integration with advanced ASIC nodes like ASAP7 highlighted significant physical-technology dependencies, the results confirm that bottleneck-driven DSE methodologies are fundamentally tool-agnostic. This study establishes a foundation for automated, ASIC-oriented HLS optimization by identifying the necessary infrastructure for cross-tool portability.

We present an MCMC-guided RRT framework that integrates Metropolis-Hastings sampling with belief-space planning and receding horizon navigation to address the challenge of efficient robot motion planning under uncertainty. Our approach replaces uniform sampling with goal-biased samples drawn from a Boltzmann distribution over a cost function that balances goal distance, obstacle clearance, and information gain, while maintaining a Gaussian distribution about the goal location. Experimental evaluation across 180 trials in maze environments of varying complexity reveals a nuanced performance profile. MCMC sampling produces significantly shorter paths when planning succeeds but fails more often than standard RRT for architectural reasons. This framework offers a principled probabilistic foundation for integrating uncertainty quantification directly into motion planning pipelines, with implications for robotic navigation in real-world environments where obstacle and goal information is incomplete.

Human gaze during manipulation is prospective — the eyes fixate the next sub-goal 1-2 seconds before the hand arrives, and return to the current object upon failure. This makes gaze a uniquely rich signal for understanding demonstrator intent. We investigate whether this signal meaningfully improves imitation learning on a bimanual ALOHA setup, collecting live ego-view gaze during teleoperation and training a UNet-based heatmap model across long-horizon tasks. Without any object labels or language supervision, our model produces interpretable attention maps that highlight task-relevant objects and implicitly track task completion as the scene evolves. We explore gaze as an attention mask, sub-task boundary detector, and autonomous sub-goal selector within a point-conditioned policy. Against expectations, a strong ACT baseline largely matches or exceeds gaze-augmented systems across all evaluated conditions, raising a core question: given sufficiently capable imitation learning baselines, where does gaze add value that policies cannot recover implicitly? We discuss failure modes, open experiments, and conditions under which gaze supervision may remain essential.

Stereotactic body radiation therapy (SBRT) is an advanced treatment for prostate cancer (PCa) that delivers high doses of targeted radiation in a short timeframe, often providing better disease control with lower toxicity. However, PCa recurrence post-SBRT remains challenging to predict using current clinical metrics. Recent advances in machine learning (ML) use radiomic features derived from multiparametric MRI (mpMRI) to improve radiation therapy (RT) response prediction. Habitat imaging generates tumor subregions that provide additional visual insights into intratumoral variation. This study investigates whether mpMRI-based habitat imaging can aid in distinguishing between PCa patients with and without recurrence 1 year post-SBRT. We hypothesize that clustering voxel-wise radiomic features can reveal distinct tumor subregions associated with these two outcome groups. Voxel-wise radiomic features were extracted from mpMRI sequences in 90 patients. Features were normalized before K-means clustering was applied to group voxels with similar features. The optimal number of clusters was determined using common evaluation metrics. Then, clusters were visualized as color-coded habitat maps overlaid on the corresponding T2w MRI. The habitat maps revealed distinctly different tumor subregions between outcome groups. Overall, this work demonstrates that habitat imaging complements radiomics-based ML models by providing insight into tumor microvasculature and may support more informed clinical decision-making in PCa treatment.