A hand-picked set of projects, each opened up for inspection rather than simply listed.
Inside every card you'll find an architecture diagram drawn from the real system, honest notes on scope and trade-offs, and links out to the code, the live deployments and the publications behind the work. Everything is sourced from the same résumé data that drives the Construct, so nothing on this page can quietly drift from the record.
You can filter which projects use which technology by toggling the tech-stack buttons below.
5 artifacts
you're looking at it
gipc.dev — self-hosted operator platform
2026
Live-demo platform on bare-metal k3s behind a Cloudflare Tunnel (zero inbound ports), infrastructure-as-code in-repo; migrated the full domain + mail estate solo.
The platform serving this page: a self-hosted operator site on bare-metal k3s behind a Cloudflare Tunnel (zero inbound ports), fully IaC + GitOps (CI → GHCR → ArgoCD). A Go core + Python/FastAPI AI service + Next.js web, with Prometheus/Grafana/Loki telemetry, a safe-by-construction infra-demo /lab, and an AI operator at /oracle.
Full-stack Next.js 15 / React 19 / Supabase product on Vercel (~15k LOC) with a three-tier test suite (Vitest, Playwright e2e, axe) and Google Calendar + email integration.
A full-stack booking product built for a real nail salon: Next.js 15 / React 19 / TypeScript on Supabase + Vercel, with Google Calendar sync and transactional email (Resend), backed by a three-tier automated test suite.
Reproducible U-Net (12.5M parameters) inverting 2,000 synthetic seismograms via a differentiable wave simulation; functional-core design, 27 deterministic tests.
A reproducible PyTorch U-Net inverting synthetic seismograms via a Deepwave differentiable wave simulation, built functional-core with deterministic tests and CUDA/MPS/CPU support.
PyTorchDeepwaveNumPyCUDA/MPS
12.5M-parameter U-Net
2,000 shot gathers via a differentiable wave simulation
L1 + SSIM loss
27 deterministic unit tests
Seismic U-Net — inversion pipeline
Reproducible PyTorch inversion — synthetic seismograms inverted through a differentiable wave simulation and a U-Net.
select a component — hover previews, click pins, Escape clears
full text: nodes, edges and facts — Seismic U-Net — inversion pipeline
Deepwave (differentiable)
Differentiable wave simulation generates the data.
seismograms (2,000 gathers)
2,000 synthetic seismograms (shot gathers).
U-Net (12.5M params)
12.5M-parameter U-Net inverts the seismograms; functional-core; CUDA / MPS / CPU.
Node.js orchestration of four fine-tuned OpenAI GPT adaptors (news-sentiment, sentiment+price, price-only and a fusion model) predicting daily S&P 500 (SPY) direction across 14 iterative fine-tuning cycles; risk-aware confusion-matrix evaluation; LaTeX manuscript.
His Master's dissertation (dissertation pathway): a Node.js orchestrator coordinating four fine-tuned OpenAI GPT adaptors over a real-world S&P 500 (SPY) dataset of 131 trading days, evaluated with a purpose-built rise/fall confusion matrix that separates missed-opportunity from capital-loss errors; results written up in a LaTeX manuscript.
AI drowning-detection research (Griffith × Huizhou)
2024–2026
First-author IEEE Access paper establishing the technical feasibility of AI drowning detection; now developing the on-device (Raspberry Pi/ARM) edge computer-vision system as the core of his PhD — targeting real-time inference — informed by a funded cross-site systematic review.
A funded Griffith × Huizhou University (China) cross-site research project: a six-database systematic review of ~82 papers with an original seven-dimension classification framework and a 13-gap research taxonomy, published first-author in IEEE Access as a feasibility study. The on-device Raspberry Pi/ARM edge-CV detection system is in-development PhD research, targeting real-time inference.