THE
STACK

Hand-written GLSL fragment shaders running at 60fps — procedural grid geometry, star field, volumetric glow, radial vignette. No Three.js. No library. Raw WebGL API against a full-screen canvas on every page.

300 frames extracted per video via ffmpeg. Loaded via chunked fetch pipeline — 8 frames at a time to avoid overwhelming the connection pool. Pre-decoded into GPU-resident ImageBitmap objects. RAF loop draws the correct frame on scroll with zero decode latency. Loading state blocks scroll until all frames are resident in GPU memory.

Drop MP4 files into finals/. Run one command. The build script extracts frames, generates page HTML from a template, rebuilds the work grid with correct thumbnails and titles, and regenerates the portfolio navigation — all automatically. Zero manual steps.

Dot tracks the mouse via requestAnimationFrame at native refresh rate — no CSS transitions, no mousemove latency. A separate ring element lerps toward the dot at 18% per frame, creating natural trailing. Expands and inverts color on any hoverable element. Collapses on click.

Overlay fires on mousedown — not click — for the fastest possible visual response. Progress bar animates to 80% immediately, completes to 100% on navigation. New page fades in from black. Every internal link on every page is intercepted by a single shared script.

LLaMA 3 8B fine-tuned via QLoRA on a merged dataset of 120K+ therapy conversation pairs from HuggingFace. DPO alignment pass using PsyCoPref preference pairs rated on empathy, safety, and autonomy. RAG layer over a notes corpus covering Ramani Durvasula, Jennifer Freyd, Sam Vaknin, Chase Hughes, Joe Navarro, and Jessica Taylor. Deployed via HuggingFace Hub to Railway.

Python pipeline that downloads therapy datasets, converts to LLaMA 3 chat template format, deduplicates, filters short and low-quality records, merges with domain-specific research corpus, and splits into train/eval sets. Runs on Kaggle GPU free tier.

Adversarial AI, benchmark design, legal reasoning, neurosymbolic learning, industrial computer vision, scaling frameworks, AI alignment, and continual learning. Each with its own designed page. Indexed on a master research page with abstracts, tags, and click-through navigation.

16 portfolio pieces, each a full-page scroll-driven animation built from After Effects renders converted to image sequences. Master work grid with thumbnail previews generated from frame 1 of each sequence. Prev/next navigation bar on every piece.

Single nav.js file injected into every page via a script tag. Hides legacy nav elements, injects the new nav bar, detects the current page for active state, skips padding-top on scroll-driven pages to preserve animation math. Portfolio-specific prev/next bar generated from a page list that rebuilds automatically with the build script.

All fonts downloaded locally. No Google Fonts CDN requests. No jQuery. No React. No framework. No build step for the frontend. Pure HTML, CSS, and vanilla JavaScript. Every file is portable, readable, and owned outright.