Primary Stack · AI Research

vision-language modelscontrastive learningexplainable AImedical imagingdiffusion & inpainting

Secondary Stack · Software Engineering

Nine years of production work, now mostly in service of shipping research.

First-author research in computer vision and multimodal ML: explainable medical imaging, representation learning for scientific data, and control of black-box generative models. Full list with abstracts and BibTeX: suxrobgm.net/research · Google Scholar

MorphoCLIP

Matches microscopy of drug-perturbed cells to plain-language descriptions of the treatment. Encoders stay frozen and only projection heads train, so it fits one consumer GPU.

CPJUMP1 · 51 plates · 3M+ images

Localize, Don't Beautify

Ask a commercial editor to change one facial feature and it beautifies the whole face. Benchmarks three localization strategies; plain masking beat prompt-only steering.

6 editors · 196 edits · identity scored

MelanomaNet

Skin lesion classification across all nine ISIC 2019 classes. GradCAM++ attention is scored against the ABCDE criteria clinicians already use, so interpretability gets a number.

ISIC 2019 · 25,331 images · 0.856 F1

Models built against real inputs rather than a clean benchmark split.

Med Image Scanner

Pulls studies straight from hospital PACS over DICOM and runs detection models over them. Predictions show up as overlays in the viewer, alongside measurement and segmentation tools. HIPAA-ready.

Bookshelf Scanner

Point a camera at a bookshelf and get back a list of what is on it. YOLO segmentation cuts out each spine, then a vision-language model reads the title and author off it.

Kept separate from the published work above.

LightDepth

Lightweight monocular depth estimation. Holds accuracy at 14.3M params where Depth Anything V2 needs 24.8M, runs 72% faster, and comes out slightly ahead on relative error on NYU Depth V2.

FSRCNN

Reproduction of FSRCNN (Dong et al., ECCV 2016) for super-resolution at 2x/3x/4x. Upsampling is learned end to end, which is where the 40x speedup over SRCNN comes from (+1.78 dB PSNR on Set5).

LogisticsX

Multi-tenant TMS for intermodal trucking. Wired into the big load boards (DAT, Truckstop), with ELD/HOS compliance, Stripe Connect, route optimization, and live tracking. DDD + CQRS architecture.

Meat.gg

60K+ users1K+ DAU

Community platform for Counter-Strike 2 servers. Profiles and messaging, a shop running on Stripe, and a native plugin that lets admins ban, report, and moderate from inside the game.

DepVault

Scans a project's dependencies across 8+ ecosystems for known vulnerabilities via OSV.dev, and doubles as an encrypted secrets vault: AES-256-GCM, one-time secret sharing, CI/CD token injection.

Blazor Form Builder

Drag-and-drop form designer that outputs JSON schema with a runtime renderer, so admin dashboards stop needing hand-written forms.

GitHub Activity

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