Rouzbeh Meshkinnejad

Designed and deployed a multi-step LLM-based agent capable of iterative retrieval, reasoning, and knowledge updates over external web sources.
The system was built for reliability, cost control, and extensibility, and integrated into internal workflows for domain experts.

Designed a production-grade Kubernetes backend for low-latency, real-time ML inference, targeting large vision transformer models. The work focused on system architecture, scalability, and operational correctness, and was formalized in a detailed design document covering infrastructure, deployment, and runtime considerations.
The system was designed around Azure Kubernetes Service (AKS) and provisioned using Infrastructure as Code (Bicep), with a strict separation between infrastructure and application layers. The architecture emphasized reproducibility, environment parity, and safe iteration, enabling model teams to deploy and update inference services independently of cluster operations.
Key design elements included autoscaling strategies for bursty inference workloads, GPU-aware scheduling, health checks and rollout policies for zero-downtime updates, and CI/CD pipelines for containerized model services. The design also addressed observability, cost control, and security boundaries between services. While the system was not deployed to production, it was reviewed and validated as a deployment-ready architecture.

Led the end-to-end design, implementation, and deployment of a large-scale image classification system for life science data, transforming a largely manual inspection workflow into an automated, ML-driven pipeline. The system scaled from a small unlabeled dataset to 2+ million high-resolution images while maintaining production-level performance and reliability.
The project began with limited labeled data, where manual annotation was prohibitively expensive. The task was to design a pipeline that could bootstrap supervision, scale efficiently, and deliver high-accuracy predictions suitable for downstream operational use. To address this, we introduced a pseudo-labeling strategy using classical image processing techniques (OpenCV) to generate initial labels from raw imagery. A custom web application was designed and deployed to visualize the pseudo-labeling process and enable domain experts to validate and refine annotations. All data and annotations were versioned and stored in Azure Blob Storage, enabling reproducible training at scale.
Model development initially relied on ResNet-based architectures, and later transitioned to DINOv2 as dataset scale and diversity increased. Training was optimized using Distributed Data Parallel (DDP) on Azure Machine Learning with mixed-precision training as well as experimenting with torch.compile backends and CUDA graphs, reducing epoch time from ~100 hours to ~7 hours on 16 T4 GPUs (≈14x speedup).
After approximately two weeks of training, the system achieved >93% accuracy on high-resolution images and >99% accuracy on small tile images. The model was deployed in a validation environment using Azure Batch, Logic Apps, and on-premises upload services, enabling seamless integration with existing workflows. This solution replaced a costly manual review process, resulting in over $100K in annual operational savings.

Enhanced a mature industrial chemical process by introducing a machine-learning layer to improve consistency and efficiency. The work focused on extracting incremental gains from large-scale operational data without altering core process mechanics.
While the existing process performed reliably, measurable variability remained across environments and operating conditions. The objective was to design a data-driven system that could identify subtle patterns in historical data and support improved outcomes at scale.
Performed extensive exploratory analysis on 10M+ tabular records, engineered features, and trained deep learning models using PyTorch. Established reproducible experimentation and model tracking with Azure Machine Learning and MLflow, and integrated the models into a semi-automated workflow suitable for global laboratory environments.
Achieved >90% predictive accuracy in production use, contributing to $500K+ in annualized savings through improved success rates and operational efficiency across deployed sites.

Redesigned and accelerated large-scale stochastic inversion workflows for geophysical signals by re-architecting computation to run efficiently on GPUs. The system preserves physically accurate forward modeling while enabling fast, practical exploration of ill-posed inverse problems.
In geophysical modeling, the forward process (i.e. observable signals such as gravity or electromagnetic responses from a known 3D subsurface property mesh) is deterministic and well-defined. The inverse problem, however, is inherently ill-posed: multiple subsurface configurations can explain the same 2D observations, requiring stochastic and iterative optimization over high-dimensional parameter spaces. Existing inversion workflows relied on legacy CPU-based software, making large-scale experimentation prohibitively slow.
I optimized the inversion pipeline by implementing GPU-native execution for the stochastic components, introducing lazy evaluation for high-dimensional tensor operations, and restructuring numerical kernels to maximize parallelism and memory efficiency without modifying the underlying physical models. Custom infrastructure supported scalable experimentation across different inversion setups while ensuring numerical stability and reproducibility.
The resulting system reduced end-to-end inversion runtime by ~12x, from ~8 hours to minutes, enabling domain experts to iterate rapidly over inversion configurations and significantly improving the practicality of advanced inversion techniques in real-world research workflows.

Developed deep learning models to predict 3D subsurface structures from 2D geophysical observations. Combined physical modeling with data-driven approaches to enhance mineral exploration accuracy.
Leveraging a hybrid of CNN-LSTM, pure LSTM, and Transformer architectures, alongside traditional ML baselines (Random Forests, XGBoost), we addressed the inverse problem of reconstructing subsurface properties from gravity, electromagnetic, and volumetric datasets. Data was generated both from the forward physical process and via FastGAN to augment limited observations. The models were trained to map 2D signals to 3D representations, enabling accurate and scalable predictions of mineral deposits. This work integrated physical insight, generative modeling, and deep learning to improve predictive performance and efficiency of real-world exploration applications.

Developed a semi-supervised deep learning pipeline to segment high-resolution satellite imagery with limited human-annotated labels. Demonstrated a scalable approach for integrating automated segmentation into geoscientific workflows.
Using ResNet-18 as the backbone, I applied a combination of semi-supervised methods—including consistency regularization and pseudo-labeling—to maximize performance under scarce annotations. Large raster images were read, tiled, segmented, and stitched into high-resolution maps, with evaluation via accuracy and Intersection-over-Union (IoU). While initially serving as a proof-of-concept, the project validated the feasibility of automated segmentation for domain experts and provided a foundation for further development toward reliable satellite image analysis.

Proposed a novel embedding-level data augmentation method for transformer-based language models, enabling flexible and controllable generation of synthetic training samples. Designed to improve generalization in low-resource NLP settings.
I introduced DoubleA, an embedding augmentation framework that operates directly in the representation space of pretrained models (BERT), addressing the limitations of word-level text augmentation. The method combines lightweight word-level perturbations with a latent-space modeling approach using SVD and Gaussian sampling to generate arbitrarily many augmented embeddings with tunable strength. DoubleA was evaluated on a downstream IMDb sentiment classification task, where it consistently outperformed standard augmentation baselines (EDA, E-Mixup, E-Stitchup) and achieved modest but reliable gains over no augmentation (≈+3% accuracy relative to other augmentation methods). This work demonstrates how controllable embedding-space augmentation can enhance robustness and generalization while remaining computationally efficient and compatible with semi-supervised learning pipelines.

Developed a hybrid modeling framework that combines historical stock price data with sentiment signals extracted from StockTwits messages to predict short-term stock price movements. The project investigated whether social-media-driven sentiment provides complementary information beyond price-only models.
The project implemented a pipeline that jointly modeled financial time-series data and sentiment derived from stock-related tweets collected from StockTwits. Sentiment features were extracted using a BERT-based classifier trained on domain-specific text, while price dynamics were modeled using neural networks operating on historical price windows and technical indicators. These components were integrated into a hybrid architecture to evaluate the impact of sentiment signals on downstream prediction tasks. Models were evaluated on real-world stock and tweet datasets using regression and classification metrics. Results showed that incorporating sentiment information led to more stable predictions and modest performance improvements under certain conditions, while also revealing challenges related to noise, timing alignment, and variability in social sentiment signals.