Full Stack Developer
Beyond shipping product, I'm drawn to the systems behind AI — search evaluation, preference-learning loops, and the math that makes ML scale in production.
I'm a full-stack developer who works at the seam between AI/ML systems and the infrastructure that keeps them honest in production — data pipelines, identity, search evaluation, and the math that makes models actually scale.
February 2025 - Present
Aquaview (IAAS) — University of Southern Mississippi
Lead the data infrastructure for a production oceanographic platform. Built a streaming statistical pipeline for 10M+ observation records, a from-scratch OAuth2/OIDC identity provider (PKCE, RS256, JS SDK) powering 15 endpoints across team apps, and a GCP Cloud Run tile server serving 268K+ COG/STAC/Zarr datasets with tuned GDAL caching and HTTP/2.
October 2025 - December 2025
ArroyoDev (Illumibot)
Shipped media-pipeline work on an AI projector platform. Profiled the segmentation engine and cut processing overhead by 40% to hit real-time on projector workloads, built a frame-accurate FFmpeg xfade transition system for multi-clip chaining, and hardened the mobile↔inference API with batching, timeouts, and retries — dropping service failures by 30%.
A CLI tool that orchestrates multiple AI coding agents (Codex + Claude Code) running in parallel. Maintains a persistent Brain knowledge base across sessions, enforces approval gates on every agentic action, runs shadow missions for strategy comparison, and manages worktree-based branch isolation per task. Built to handle the coordination problem that emerges when more than one agent can touch a codebase.
Autonomous overnight ML research pipeline. Claude proposes protein interaction modifications, Modal dispatches experiments to H200 GPUs, a 5-seed Welch's t-test validates each result, and a keep/revert state machine decides whether to commit the change — all without human intervention. Runs ~12 experiments/hour on infrastructure spanning Modal, Colab, FastAPI, a Gradio dashboard, a full CLI, and 38 tests.
An adversarial benchmarking framework for neural search engines (Exa, Tavily, Brave, SerpAPI). 13 embedding-theory-grounded attack categories, 70+ adversarial test pairs, LLM-as-judge scoring, and statistical rigor (bootstrap CIs, Benjamini-Hochberg correction) — surfacing where production search actually fails on negation and numeric queries.
A never-ending preference-learning loop for LLMs: feedback collection → semantic deduplication → SFT/DPO adapter distillation → inference server. Designed to update a model hourly without full retrains, treating model behavior like a product that ships every sprint instead of every quarter.
A browser-native Sinkhorn solver running ε-regularized optimal transport with log-domain stabilization and ε-annealing — entirely in the browser via WebGPU. Handles 512×512 grids (262,144 points) at interactive speeds, 3–8× faster than the naive path and stable at very small ε.
Let's discuss your project and explore how we can create something extraordinary together.