By operating in the diffraction-limited regime, astrophotonic instruments break the scaling relationship between telescope and instrument sizes, unlocking increasingly high-resolution spectrographs in a compact, stable, and inexpensive format. Here, we present the development of multiple integrated software systems for on-sky tests of a chip-based arrayed waveguide grating (AWG) spectrograph using the Hale telescope at Palomar Observatory. The spectrograph is sensitive to the NIR H-band (1.45-1.65 μm) with a spectral resolution (λ/Δλ) of about 30,000. This assembly integrates a photonic spectrograph, physical bulk optical components, motorized control units, calibration lasers, and an InGaAs infrared detector packaged into a fiber-fed instrument within a physical footprint equivalent to a shoebox. The package is made complete with customized imaging and control software and a data-reduction pipeline that delivers wavelength-calibrated spectra. This pathfinder performance demonstration paves the way to development of mature facility-class photonic instruments for ground- and space-based observatories.

Massive binary stars located near the centers of galaxies can undergo dramatic evolution under the gravitational influence of a central supermassive black hole. These systems form hierarchical triples, where the outer black hole can drive long-term orbital changes in the inner binary through secular dynamics. Existing work often isolates dynamical interactions, neglecting detailed stellar evolution, or uses simplified prescriptions that overlook key phases such as mass transfer, common-envelope evolution, and compact-object formation. This project asks how combining realistic stellar evolution with detailed triple dynamics changes the expected evolution of these systems. To address this question, we use computational simulations that couple modern stellar evolution prescriptions with dynamical treatments of hierarchical triples. This approach allows us to study how binaries evolve in environments where gravitational perturbations from a supermassive black hole can strongly reshape orbital properties over time. Our results show that the interplay between dynamics and stellar evolution produces a wider variety of outcomes than predicted by simplified approaches, including mergers and binary evaporation. By bridging two previously isolated or otherwise simplified approaches, this work provides a more realistic framework for understanding stellar populations and transient phenomena in the extreme environments at the centers of galaxies.

Planetary surface processes shape the morphology and distribution of regolith on Mars, and quantitative clast analysis can further our understanding of these processes. This project investigates whether machine learning based segmentation can enable robust measurement and interpretation of Martian regolith structure across rover sols. By training and using machine learning models building on prior enhancements to the Segment Anything Model and Mask R-CNN, I apply a custom segmentation pipeline to Mastcam-Z and WATSON imagery to generate high fidelity clast masks. Using Python-based workflows, I extract geometric features including clast area, long-axis length, and short-axis length derived from polygon masks. Measurements are converted from pixel space to physical units and aggregated across sols to analyze spatial and temporal variability. Sol-indexed statistics reveal substantial variability in clast morphology, with early sols exhibiting significantly larger clasts and later sols showing smaller characteristic sizes and greater dispersion. Population-level size-frequency trends further indicate a shift from coarse, boulder-dominated surfaces to finer regolith distributions over time. These results motion towards a future of glorified results from better and strictly trained ML models to clast said images.

Large language Models (LLMs) are increasingly deployed in settings involving complex interactions among humans and AI agents. As these systems are used to model social behavior and collective decision-making, understanding how they internally represent identity, belief, and bias becomes critical. Our research investigates how identity and semantic opinion are encoded and evolve within the latent states of large language model agents during social interaction. To model social networks, we initialized agents represented by Llama 3.2 with controlled personas and continuous opinion values, prompted through simulated interactions, and analyzed via hidden-state embeddings extracted throughout the transformer. Latent representations are examined using alignment methods, PCA trajectory visualization, and layerwise trajectory comparison to characterize how agent representations evolve across contexts and model depth. We leverage supervised probe models to recover assigned opinion values from hidden states and compare them against token-based sentiment baselines. Latent-state probes outperform text-based estimation, indicating that internal representations preserve semantic opinion more faithfully than surface outputs. Additionally, agents with similar opinions but different personas remain close in latent space until later transformer layers, where stylistic divergence emerges. These findings suggest latent representations provide an interpretable framework for modeling belief and identity dynamics in LLM agent systems.