Shrey Shah

Nishi Doshi, Shrey Shah

Hospitals run more machine learning on GPUs while the carbon footprint of grid electricity rises and falls through the day. Using a computer simulation, we compare $13$ scheduling rules on mixed GPU hardware, with synthetic patient-style jobs, urgency tiers, and time-of-day carbon traces. We do not study patient outcomes; every percentage we report is a simulator queue number, not a clinical finding. We ask whether running non-urgent jobs overnight is almost as good as a richer rule that mixes urgency and carbon (CUCA at weight 0.45, written CUCA$_{0.45}$). The comparison keeps carbon reduction secondary to clinical priority and deadline compliance, so each policy is judged on both average kg CO$_2$e and missed-deadline behavior. CarbonGreedy and CarbonShift are carbon-first stress tests that demonstrate how poorly wrong vendor presets can disrupt clinical priorities, and are not meant for production. Numbers are averages over many test settings, with wide run-to-run spread and no statistical adjustment, so headline ratios are exploratory. On an eight-GPU baseline, the overnight rule closes about $78\%$ of the carbon gap between urgency-only and CUCA$_{0.45}$ while missing fewer urgent deadlines than either. CarbonShift lets about $46\%$ of the most urgent jobs miss their deadline; this is simulated queueing, not bedside harm. At $48$ jobs per hour, the carbon footprints almost tie, yet the overnight rule still misses fewer urgent deadlines. A geography test, where regions share one daily carbon shape with only timezone shifts, trims under one percentage point of average carbon; a twelve-hour routine window saves a little carbon for CUCA$_{0.45}$ but raises overall missed deadlines. Overnight batching stays competitive on average modelled carbon; carbon-only rules belong only in stress tests.

Katrina Ashton, Chahyon Ku, Shrey Shah et al.

Language-specified mobile manipulation tasks in novel environments simultaneously face challenges interacting with a scene which is only partially observed, grounding semantic information from language instructions to the partially observed scene, and actively updating knowledge of the scene with new observations. To address these challenges, we propose HELIOS, a hierarchical scene representation and associated search objective. We construct 2D maps containing the relevant semantic and occupancy information for navigation while simultaneously actively constructing 3D Gaussian representations of task-relevant objects. We fuse observations across this multi-layered representation while explicitly modeling the multi-view consistency of the detections of each object using the Dirichlet distribution. Planning is formulated as a search problem over our hierarchical representation. We formulate an objective that jointly considers (i) exploration of unobserved or uncertain regions of the environment and (ii) information gathering from additional observations of candidate objects. This objective integrates frontier-based exploration with the expected information gain associated with improving semantic consistency of object detections. We evaluate HELIOS on the OVMM benchmark in the Habitat simulator, a pick and place benchmark in which perception is challenging due to large and complex scenes with comparatively small target objects. HELIOS achieves state-of-the-art results on OVMM. We demonstrate HELIOS performing language specified pick and place in a real world office environment on a Spot robot. Our method leverages pretrained VLMs to achieve these results in simulation and the real world without any task specific training.

Nishi Doshi, Shrey Shah

Diabetic Retinopathy (DR) is one of the leading causes of preventable blindness, yet rural regions often lack the specialists and infrastructure needed for early detection. Although cloud-based deep learning systems offer high accuracy, they face significant challenges in these settings due to high latency, limited bandwidth, and high data transmission costs. To address these challenges, we propose a two-tier edge-cloud cascade on the public APTOS 2019 Blindness Detection dataset. Tier 1 runs a lightweight MobileNetV3-small model on a local clinic device to perform a binary triage between Referable DR (Classes 2-4) and Non-referable DR (Classes 0-1). Tier 2 runs a RETFoundDINOv2 model in the cloud for ordinal severity grading, but only on the subset of images flagged as referable by Tier 1. On a stratified APTOS test split of 733 images, Tier 1 reaches 98.99% sensitivity and 84.37% specificity at a validation-tuned high-sensitivity threshold. The default cascade forwards 49.52% of test images to Tier 2, reducing cloud calls by 50.48% relative to using a cloud-based model for all images. In the deployed 4-class output space (Class 0-1 / Class 2 / Class 3 / Class 4), the cascade obtains 80.49% accuracy and 0.8167 quadratic weighted kappa; the cloud-only baseline obtains 80.76% accuracy and 0.8184 quadratic weighted kappa. On APTOS, the cascade cuts cloud use by about half with a modest drop in grading performance. Index Terms: Diabetic Retinopathy, Edge-Cloud Cascade, MobileNetV3-small, RETFound-DINOv2, Retinal Screening, tele-ophthalmology

Shrey Shah, Justin Wagle

Multi-adapter serving systems route entire sequences to a single adapter, forcing a choice when requests span multiple domains. This assumption fails in two important settings: (1) multimodal generation, where text and image tokens require different adapters within the same sequence, and (2) mixed-capability requests like "write code to solve this equation," which need expertise from multiple specialized adapters. We introduce per-token routing, which routes individual tokens to adapters based on either vocabulary structure (for multimodal models) or learned gating (for semantic specialization). Per-token routing is provably optimal, achieving work N for N tokens versus K \cdot N for per-sequence routing with K adapter types. Our key contribution is MoLoRA (Mixture of LoRA), which enables composable specialization: load multiple domain-specific adapters and let a learned router select the appropriate adapter per-token. We demonstrate that specialization dramatically beats scale: MoLoRA enables Qwen3-1.7B to exceed Qwen3-8B across four reasoning benchmarks while being 4.7x smaller. This enables modular expertise at inference time: train focused LoRAs independently, combine them without retraining, and add new capabilities by simply loading new adapters.

Aditya Pratap Singh, Shrey Shah, Ramanakumar Sankar et al.

Insights into Jupiter's atmospheric dynamics are vital for understanding planetary meteorology and exoplanetary gas giant atmospheres. To study these dynamics, we require high-resolution, photometrically calibrated observations. Over the last 9 years, the Juno spacecraft's optical camera, JunoCam, has generated a unique dataset with high spatial resolution, wide coverage during perijove passes, and a long baseline. However, JunoCam lacks absolute photometric calibration, hindering quantitative analysis of the Jovian atmosphere. Using observations from the Hubble Space Telescope (HST) as a proxy for a calibrated sensor, we present a novel method for performing unpaired image-to-image translation (I2I) between JunoCam and HST, focusing on addressing the resolution discrepancy between the two sensors. Our structure-preserving I2I method, SP-I2I, incorporates explicit frequency-space constraints designed to preserve high-frequency features ensuring the retention of fine, small-scale spatial structures - essential for studying Jupiter's atmosphere. We demonstrate that state-of-the-art unpaired image-to-image translation methods are inadequate to address this problem, and, importantly, we show the broader impact of our proposed solution on relevant remote sensing data for the pansharpening task.

Shrey Shah, Alex Lin, Scott Lin et al.

Accurate damage prediction is crucial for disaster preparedness and response strategies, particularly given the frequent earthquakes in Turkey. Utilizing datasets on earthquake data, infrastructural quality metrics, and contemporary socioeconomic factors, we tested various machine-learning architectures to forecast death tolls and fatalities per affected population. Our findings indicate that the Random Forest model provides the most reliable predictions. The model highlights earthquake magnitude and building stability as the primary determinants of damage. This research contributes to the reduction of fatalities in future seismic events in Turkey.

Justin Kiefel, Shrey Shah

In this work, we introduce LokiLM, a 1.4B parameter large language model trained on 500B tokens. Our model performs strongly in natural language reasoning tasks and achieves state-of-the-art performance among models with 1.5B parameters or less. LokiLM is trained using multi-teacher knowledge distillation and high-quality training data to achieve benchmark results competitive with larger models trained on significantly more tokens. We support these findings by introducing steps to avoid benchmark contamination and overfitting throughout our development process. Despite its promising performance, LokiLM exhibits a concerning amount of hallucinations and scores poorly on the TruthfulQA benchmark, so we do not release the model publicly.

Shreyash Mishra, Shrey Shah, Rex Pereira

This paper introduces the Multi-Step Action Model (MSAM), a closed-source AI model designed by Empsing to address challenges hindering AI adoption in enterprises. Through a holistic examination, this paper explores MSAM's foundational principles, design architecture, and future trajectory. It evaluates MSAM's performance via rigorous testing methodologies and envisions its potential impact on advancing AI adoption within organizations.

Scott Sievert, Shrey Shah

Mini-batch stochastic gradient descent (SGD) and variants thereof approximate the objective function's gradient with a small number of training examples, aka the batch size. Small batch sizes require little computation for each model update but can yield high-variance gradient estimates, which poses some challenges for optimization. Conversely, large batches require more computation but can yield higher precision gradient estimates. This work presents a method to adapt the batch size to the model's training loss. For various function classes, we show that our method requires the same order of model updates as gradient descent while requiring the same order of gradient computations as SGD. This method requires evaluating the model's loss on the entire dataset every model update. However, the required computation is greatly reduced by approximating the training loss. We provide experiments that illustrate our methods require fewer model updates without increasing the total amount of computation.