Welcome to Vice City
Software Engineer · ML & Computer Vision Researcher
EmTech Care Labs | Portland, ME
Lead engineer on an Alzheimer's care platform on AWS, where 200+ caregivers and patients message each other and share care plans. I folded three web and mobile codebases into one TypeScript monorepo and did the security work (encryption, audit logs, access control) that got us through an outside HIPAA readiness review with no critical findings. The ML features are mine too: a decline-risk model and an NLP pipeline for clinical notes. I also mentor two junior engineers.
AllFactors | Santa Clara, CA
Moved an old ASP.NET Web Forms real-estate analytics app to Blazor WebAssembly, and pages got about twice as fast. I built the live pricing feed (SignalR plus SQL Server change tracking) that pushed updates to hundreds of users in under a second, along with the component library all the dashboards were built from. The valuation and forecast models behind those dashboards retrained every night on the day's transactions.
Smart Meal Service | Russia
Wrote the C# software that ran a self-service food kiosk and a robotic cashier, payment terminal and hardware protocols included. I also built the customer ordering app (Xamarin) and the admin dashboard (ASP.NET Core and Angular). A hardware abstraction layer let the same code run on all three machine models, and orders went to the right kitchen station automatically.
Pentalight Technology | Malaysia
Built multiplayer networking, spatial audio, hand tracking, and the in-world UI for a VR smart-city simulation in Unity, plus a desktop client for anyone without a headset. Scene builds used to take hours of hands-on work. After I automated the art pipeline (model import, LODs, light baking), an artist could start one and walk away.
Matches microscope images of treated cells to a plain-language description of the drug or gene behind the change. The DINOv3 and BioClinical ModernBERT encoders stay frozen and only small heads are trained, so one consumer GPU is enough. Tested on CPJUMP1 (51 plates, 3M+ images).
Image editors asked to change one facial feature tend to retouch the whole face, which ruins a surgical preview. I tried three client-side fixes across six commercial editors and 196 edits, checking identity with ArcFace and whether each edit stayed put. Plain masked compositing worked best.
A nine-class skin-lesion classifier (EfficientNet V2, focal loss) with 0.86 weighted F1 on 25K ISIC 2019 images. Its GradCAM++ heatmaps are broken down by the ABCDE criteria dermatologists already use, so a clinician can see what the model was looking at.
Estimates depth from one image with a ResNet18 encoder and U-Net decoder. Next to Depth Anything V2 it has 42% fewer parameters (14.3M vs. 24.8M), runs 72% faster, and makes smaller errors on NYU Depth V2.
Rebuilt FSRCNN (Dong et al., ECCV 2016) from the paper at 2x, 3x, and 4x. My version reproduced its gains over SRCNN on Set5 (+1.78 dB PSNR) and Set14 (+1.26 dB), and I added ablations on the shrinking and mapping layers.
LLM dispatch agent, 4 portals, live app
Software for running a trucking company. A multi-provider LLM agent with custom tools matches loads to trucks, checks drivers' federal hours-of-service limits, and plans multi-stop routes. The rest of the system is four web portals, a Kotlin Multiplatform driver app, a multi-tenant .NET backend (DDD/CQRS), Stripe payments, and integrations with DAT, Truckstop, Samsara, and Motive.
Runs on your Claude Code or Codex plan
An AI agent that applies to jobs for you. It searches LinkedIn, Indeed, and other boards, scores each posting against your resume, tailors the resume, and fills in the application in a real browser on your own machine. It runs inside Claude Code or Codex, so there's no API key or usage bill. A web dashboard tracks every application from applied to offer, and a small .NET host links the agent to it.
50K+ registered players
Community site for Counter-Strike 2 servers, with 50K+ registered players and over 1K active on a typical day. Players get profiles, friends, messaging, and a cosmetics shop (Stripe). A C++ Metamod:Source 2 plugin ties the game servers to the backend, so admins can ban, handle reports, and moderate without leaving the game.
8 package ecosystems scanned
Checks your dependencies against OSV.dev for known vulnerabilities in eight ecosystems, including npm, PyPI, NuGet, Cargo, and Maven, and gets through 1,000+ of them in seconds. It's also an encrypted secrets vault (AES-256-GCM) with one-time sharing and a .NET AOT CLI that injects tokens into CI/CD pipelines.
X-ray, CT and MRI from hospital PACS
Pulls X-ray, CT, and MRI studies from hospital PACS over DICOM and runs PyTorch detectors on them, for example pneumonia on chest X-rays and brain bleeds on head CTs. Results show up as overlays in the OHIF viewer. Scans are de-identified on the fly, and access is role-based and logged. Built with FastAPI and Next.js.
Outstanding Project Award, Northeastern
Snap a photo of a bookshelf and get back a list of the books. YOLO picks out each spine, Moondream2 (via llama.cpp) reads the title and author, and an Angular UI lets you fix mistakes and export. It won an Outstanding Project Award at Northeastern.