Kevin Yoder

Detects and FDI-numbers teeth, segments them, and flags pathology, then an LLM writes the insurance narrative grounded in those findings. A two-phase arc from a segmentation research effort to a shipped CV-grounded pipeline.

An EfficientNet-B0 that labels a radiograph bitewing vs periapical — 99.6% on 550 unseen external images — shipped as a torch-free ONNX artifact three systems consume from one file.

Ported a night-guard CAD add-on from Blender 2.79 to 4.5 — 166 operators, a bgl→gpu rewrite, an OpenVDB port — and deleted ~220,000 lines of dead code after tracing the import graph.

Intraoral smile and whitening simulation plus clinical analysis — face-parsing and SDXL inpainting via ComfyUI, with a quality gate the clinician never sees fail.

Turns a patient's full chart into a concise clinical summary with mandatory pre-med and severity flags. De-identified inputs only, with a three-layer PHI defense — described honestly, not badged "HIPAA-safe."

A zero-dependency, single-file tool to get phone photos onto the operatory PC's imaging system — scan a QR, tap-select, done. No cables, no cloud; photos never leave the practice LAN.

Hardening the homelab across host, network, container, and app layers: key-only SSH + a hand-rolled sysctl suite, default-deny UFW, Tailscale-only ingress, a multi-agent audit of 19 services, and an in-progress pass that caught Docker quietly bypassing the firewall.

Provider-agnostic create → train → export → destroy, with a strict teardown invariant. Cost-engineered to ~$0.10–$2 per run (~$9 for a full 5-cycle detector campaign) via spot economics and provider arbitrage.

A self-run bake-off of 14 coding-agent CLIs on one fixed "build + deploy this app" task, auto-scored on infra checks, rule compliance, and LLM-as-judge quality — 1,152 possible runs, six weekly cron matrices.

QLoRA fine-tunes of Llama 3.2/3.1 (SFT + DPO) → GGUF → Ollama, fed by an automated, secret-sterilized ops-history data pipeline. The local model scored 100% vs a 5% base on homelab facts.

One prompt → text, image, music, speech, or 3D, over a local model fleet with local GPU offload and a custom ComfyUI node pack — a coherent toolkit, not scattered apps.

Reusable skills that extend an AI agent with disciplined workflows — a high-fidelity save-state protocol and a rigorous legacy-code reverse-engineering routine with mandatory file:line citations. Building capabilities for agents.

A game-asset aggregator with 20 source connectors (Kenney, OpenGameArt, Poly Haven, Freesound, NASA…) plus a dedupe / thumbnail / AI-tag pipeline. The supply line behind the game platform.

Surfaces only what's streamable on the household's subscriptions across 9 services, composite-scored, with a Family / adults / Smart (Claude + Gemini ensemble) mode split and Roku LAN control. 201 tests. In use.

A "Tinder for food": a child joins by room code and swipes food cards until N likes reveal the meal, paired with an LLM self-expanding catalog and a three-tier image pipeline. Both services tested.

PowerShell + UniFi local-API tooling that provisions an isolated IoT VLAN, an SSID, and a nine-rule zone-based firewall on a UDM, with a Terraform-style plan / apply / verify / destroy workflow.

Downloads one-off videos or whole playlists and auto-imports them into Jellyfin by channel — and you can subscribe to channels for a personal YouTube PVR.

A browser remote with real hardware control — SSDP discovery, Roku ECP over the LAN, and DHCP-resilient tracking that re-resolves a device by its hardware serial when its IP changes.

~115 lines: reads the home library and asks an LLM for three movies and two shows for tonight, each with a one-line reason, cached and refreshed every six hours. The judgment, not the size.

A family AI storytelling pair: generate an illustrated-length story with an LLM, then have it read aloud paragraph-by-paragraph in a neural voice — with a 3-ahead prefetch, resume-from-failed-chunk read-aloud engine across ~54 Kokoro + ~30 Gemini voices.