Primary Stack  ·  AI Research

vision-language models   contrastive learning   explainable AI   medical imaging   diffusion & inpainting

Secondary Stack  ·  Software Engineering

Nine years of production work. These days I mostly use it to ship research.

I do computer vision research, mostly for medicine and biology. Abstracts and BibTeX are on suxrobgm.net/research, citations on Google Scholar.

MorphoCLIP

Matches microscope images of treated cells to a plain-language description of the drug or gene behind the change. The encoders stay frozen and only small heads train, so one consumer GPU is enough.

CPJUMP1 · 51 plates · 3M+ images

Localize, Don't Beautify

Commercial image editors asked to change one facial feature tend to retouch the whole face. I compared three ways to keep the edit local, and plain masking did better than prompt-only steering.

6 editors · 196 edits · identity scored

MelanomaNet

Sorts skin lesions into all nine ISIC 2019 classes, then checks the model's GradCAM++ attention against the ABCDE criteria dermatologists already use and scores how well the two agree.

ISIC 2019 · 25K images · 0.86 F1

Med Image Scanner

Pulls studies from a hospital's PACS over DICOM and runs detection models on them. Predictions appear as overlays in the viewer next to the usual measurement and segmentation tools, and the whole thing is built for HIPAA.

Bookshelf Scanner

Snap a photo of a bookshelf and get back a list of the books. YOLO picks out each spine and a vision-language model reads the title and author.

LightDepth

Depth from a single image in 14.3M parameters, against 24.8M for Depth Anything V2. It runs 72% faster and has slightly lower relative error on NYU Depth V2.

FSRCNN

My reimplementation of FSRCNN (Dong et al., ECCV 2016) for 2x/3x/4x super-resolution. Learning the upsampling end to end is what makes it 40x faster than SRCNN, and it still gains +1.78 dB PSNR on Set5.

LogisticsX

Multi-tenant system for running an intermodal trucking company. A multi-provider LLM agent handles dispatch, and the platform plugs into DAT and Truckstop, tracks ELD hours-of-service, plans routes, shows trucks live, and handles payouts through Stripe Connect. Built on DDD and CQRS.

Meat.gg

60K+ users   1K+ DAU

Community site for Counter-Strike 2 servers, with profiles, messaging, and a Stripe shop. A native server plugin lets admins ban players, handle reports, and moderate without leaving the game.

DepVault

Checks a project's dependencies against OSV.dev for known vulnerabilities across 8+ ecosystems. It's also an encrypted secrets vault (AES-256-GCM) with one-time sharing and a CLI that injects tokens into CI/CD.

Blazor Form Builder

Drag-and-drop form designer for Blazor. Forms are saved as JSON schema and rendered at runtime, so admin dashboards don't need hand-written forms.

GitHub Activity

All four regenerate daily from the GitHub GraphQL API via activity_chart.py.