Roman Girin

Senior Modeling Software .NET Developer / AI Engineer

AI & LLM Engineering

• AI Orchestration into .NET Ecosystem: architected and deployed production enterprise model orchestration and agentic workflows with RAG, leveraging vector embeddings (pgvector). Built an MS Semantic Kernel plugin to integrate core software via ASP.NET 8 REST HATEOAS API, enabling tooling integration into the LLM assistant. Monitored the AI assistant in production via Lunary.
• Python-Driven LLM R&D: conducted target R&D and executed open-source LLM fine-tuning via LoRA using the Transformers library (Python, PyTorch) to customize, optimize, and align open-source models for domain-specific enterprise workflows. Post-fine-tuning model evaluation via EleutherAI.
• R&D in Mathematical Optimization (Graph ML): designed and built a proof-of-concept lightweight Graph Neural Network (GCNN) to estimate LP (Linear Programming) task feasibility pre-solve, reaching 92% accuracy. Mapped LP matrix representations to bipartite graphs using PyTorch Geometric (PyG) — see the GCNN write-up.
• Team Upskilling: introduced agentic AI workflows (Claude Code) across the engineering team, raising development velocity.

Software Architecture & .NET Development

• Architecture Modernization: spearheaded the architectural transition from a legacy 2-tier to 3-tier architecture. Migrated the Business Logic Layer (BLL) from the database tier (MSSQL T-SQL) into the .NET application layer (C#), improving system maintainability and resulting in a 30% reduction in support ticket volume. Implemented compatibility with PostgreSQL.
• Windows to Linux Migration: spearheaded the core codebase migration from .NET Framework to cross-platform .NET 8, implementing a modern asynchronous programming model (async/await) that slashed system resource utilization by 15%.
• Microservices in Cloud: decomposed an on-premise monolith into high-load microservices deployed on Azure Container Apps and Azure Kubernetes Service (AKS). Designed inter-service communication using REST and asynchronous messaging EDA (Azure Event Grid / Queues), ensuring distributed data consistency via Saga, CQRS, and Outbox patterns.
• Native Interoperability: developed critical components in C++ to seamlessly integrate CLR-managed and native code blocks — see the interop bridge write-up.
• Engineering Culture: practice DDD, SOLID, GoF. Established testing standards by implementing comprehensive unit and integration testing (xUnit, NSubstitute) and Test-Driven Development (TDD). Managed build and deployment workflows via Azure DevOps pipelines.

Senior Computer Vision Engineer (Part-Time)

ANPR System Development

• System Design: designed and implemented an Automatic Number-Plate Recognition (ANPR) system driven by custom deep networks (MTCNN detector, STR-Net recognizer).
• CNN Architecture Optimization: conducted deep structural analysis of MTCNN and re-engineered the detection pipeline around P-Net + O-Net, eliminating R-Net entirely — lower GPU load with zero accuracy degradation.
• Real-Time Processing: implemented the core video-frame processing loop in C++, reaching 11–13.5 ms (~75–90 fps) per frame — about 1.5× faster than the Python implementation.
• Object Tracking: built and tuned a vehicle-tracking module on DeepSORT/BoT-SORT/StrongSORT, fine-tuning OSNet (Omni-Scale Network) for vehicle re-identification embeddings — 72% mAP, 94% Rank-1.
• Data Management: customized LabelImg to the company's metadata format, led an annotation team, built a dataset-validation pipeline, and set up on-prem data/experiment versioning (MLFlow, DVC).
• Business Impact: deployed across multiple sites — ALPR gives advance notice of road trains approaching unloading warehouses, eliminating delays entirely; also drives corporate-fleet tracking and automated gate access control at residential and industrial sites.

CCTV Compliance-Audit System Development

• Detection System Development: built and trained an SSD (Single Shot MultiBox Detector)-based object detector to track specialized commercial cleaning vehicles in real-time video streams.
• Data Engineering: automated data collection, filtering, and image preprocessing tuned for non-standard, changing indoor lighting conditions.
• Business Impact: deployed across "Lastochka" express-train maintenance depots, eliminating cleaning-schedule violations and entirely preventing regulatory compliance penalties.

Computer Vision Engineer (Part-Time / Fellowship)

• System Design: engineered a machine vision diagnostic system utilizing Deep Convolutional Neural Networks (CNNs) to automate the detection of broken railway fasteners and bindings, using Python, PyTorch, OpenCV, TorchVision, and CUDA/cuDNN acceleration.
• Dataset & MLOps: gathered imaging dataset for CNN training, successfully managed severe data class imbalances by implementing synthetic generation and augmentation techniques via imbalanced-learn (SMOTE).
• Optimization & Edge Deployment: converted models to TensorRT and ONNX formats using FP16/INT8 quantization, optimizing for real-time inference on edge computing platforms (NVIDIA Jetson Nano and AGX Xavier).
• Integration & Async Backend: tied the imaging dataflow into enterprise infrastructure via a custom REST HATEOAS API (Django REST Framework) and asynchronous messaging streams (Kafka/RabbitMQ).
• Camera Calibration: experienced in configuring Ethernet cameras for CV systems, specifically working with scientific Baumer cameras and their neoAPI.
• Business Impact: enabling Jetson-powered inference inside mobile laboratory trains at JSC RPC "InfoTrans" allowed defect discovery at full transit speed, boosting defect detection rates by 60%.

Senior Software Specialist (Promoted for Outstanding Performance)

• Outstaffing at Sberbank CIB: developed backend functionality on the Diasoft platform for the investment banking business.
• High-Load Databases: designed, developed, and optimized complex stored procedures, ER-models and data-tier components for high-load transaction systems across Sybase ASE and MSSQL database engines.
• Technical Estimation: conducted technical analysis and pre-sale engineering evaluations for resource and cost estimation.

Project Manager / .NET Developer

• Systems Analysis & Design: partnered with clients to formalize requirements into detailed Technical Specification Documents (TSDs) for large-scale business process automation.
• Full-Stack C# Engineering: managed end-to-end Enterprise Content Management (ECM) implementations while actively writing custom C# extensions and core integration modules.