Welcome to Vice City
Senior Full Stack Developer | AI & Computer Vision Researcher

I write software and I do computer vision research. Nine years of the first, and a Master's in Computer Science from Northeastern for the second. The work I like best is taking a model that only runs in a notebook and turning it into something people use.
My research is in computer vision and deep learning, mostly medical imaging and explainable AI, with three papers on arXiv. Along the way I have built systems that classify skin lesions, match microscopy images to text, estimate depth from a single photo, and read the titles off a shelf of books.
Lead engineer on a HIPAA-compliant Alzheimer's care platform on AWS: real-time messaging, push notifications, and shared care plans for 200+ caregivers and patients. Consolidated three web and mobile codebases into one TypeScript monorepo with a shared component library. Encrypted PHI, audit logging, and role-based access took the platform through an outside HIPAA readiness review with no critical findings. Also built its ML features, a decline-risk model over care-log data and an NLP pipeline that structures clinical notes, and mentored two junior engineers.
Rebuilt a legacy ASP.NET Web Forms real-estate analytics app on Blazor WebAssembly, with page loads about twice as fast. Built the real-time market-data pipeline on SignalR and SQL Server change tracking that streams live pricing to hundreds of concurrent users in under a second, and the shared Blazor component library every dashboard is built on. Property-valuation and market-forecast models retrain nightly on the day's transactions and serve into those same dashboards.
C# / .NET control software for a self-service food kiosk and a robotic cashier, down to the payment-terminal and hardware protocols. Customer ordering app in Xamarin, admin dashboard in ASP.NET Core and Angular. A hardware abstraction layer lets one codebase run on all three machine models, and automated order routing sends each order straight to the right kitchen station.
Multiplayer networking, spatial audio, hand tracking, and in-world interfaces for a VR smart-city simulation in Unity, plus a desktop spectator client for people without a headset. Automated the art pipeline (model import, LOD generation, lighting bakes), so a multi-hour job per scene became one an artist could start and walk away from.
"Sukhrob is one of the most versatile developers I've worked with. He took a vague product idea and turned it into a polished, production-ready application in just 8 weeks. His ability to handle everything from database design to pixel-perfect UI is genuinely impressive. Already planning our next project together."
"We brought Sukhrob on to rescue a project that was falling behind. He quickly identified the architectural bottlenecks, refactored the core systems, and delivered a stable release two weeks ahead of our revised deadline. Clear communicator, great problem solver, zero drama."
"Finding a developer who understands both game networking and web infrastructure is nearly impossible — Sukhrob is that rare find. He built our entire multiplayer backend and admin dashboard from scratch. The codebase he left us is clean, well-documented, and easy to maintain."
Contrastive image-text model that matches Cell Painting images of drug- and gene-perturbed cells to plain-language treatment descriptions. DINOv3 and BioClinical ModernBERT backbones stay frozen with small projection heads trained on top, so it trains on one consumer GPU. Evaluated on CPJUMP1: 51 plates, 3M+ images.
Ask a commercial image editor to change one feature of a face and it beautifies the whole thing, a problem when the picture is a surgical preview. Compared prompt steering, masked compositing, and model-based inpainting across six editors on 196 edits, scored with ArcFace for identity and a localization metric. Plain masked compositing localized best.
EfficientNet V2 classifier with focal loss over all nine ISIC 2019 categories, 0.86 weighted F1 on 25K images. GradCAM++ heatmaps are broken down along the ABCDE criteria dermatologists already use, so a clinician can check what the model looked at.
ResNet18 encoder with a U-Net decoder for monocular depth. 42% fewer parameters than Depth Anything V2 (14.3M against 24.8M), 72% faster inference, and lower relative error on NYU Depth V2.
Reproduction of FSRCNN (Dong et al., ECCV 2016) at 2x, 3x, and 4x with end-to-end learned upsampling. The paper's gains over SRCNN held up on Set5 (+1.78 dB PSNR) and Set14 (+1.26 dB), with added ablations on the shrinking and mapping layers.

Transportation management system for trucking companies. A Claude tool-use agent matches loads to trucks, checks federal hours-of-service, and plans multi-stop routes. Around it: four web portals, a Kotlin Multiplatform driver app, a multi-tenant .NET backend on DDD/CQRS, Stripe payments, and integrations with DAT, Truckstop, Samsara, and Motive.

Community platform for Counter-Strike 2 servers, 50K+ registered users and 1K+ daily actives: profiles, messaging, friends, and a cosmetics shop on Stripe. A native C++ Metamod:Source 2 plugin links the live game servers to the web backend, so admins ban, report, and moderate from inside the game.

Dependency vulnerability scanner across eight language ecosystems (npm, PyPI, NuGet, Cargo, Maven, and more), checking each against OSV.dev; 1,000+ dependencies in seconds. Doubles as an encrypted secrets vault (AES-256-GCM) with one-time sharing and a .NET AOT CLI that injects tokens into CI/CD pipelines.

Pulls X-ray, CT, and MRI studies from hospital PACS over DICOM and runs PyTorch detectors: pneumonia on chest X-ray, intracranial hemorrhage on head CT. Predictions render as overlays inside OHIF, with on-the-fly de-identification, audit logging, and role-based access. FastAPI backend, Next.js frontend.

An online chess platform with both human and AI opponents, rated or casual. Moves go over WebSockets, so both boards stay in sync as the game runs.
Point a camera at a bookshelf and get back a list of what is on it. YOLO instance segmentation isolates each spine, then Moondream2 (via llama.cpp) reads the title and author. FastAPI backend, Angular UI for corrections and export. Course Outstanding Project Award.
Northeastern University, Boston, MA
January 2024 - May 2026
GPA: 4.0/4.0Suffolk University, Boston, MA
September 2021 - May 2023
Cum Laude | GPA: 3.5/4.0INTI International College, Malaysia
September 2019 - July 2021
TUIT, Uzbekistan
September 2017 - June 2019
GitHub
github.com/suxrobgmTelegram
@suxrobgm