Hi, I'm Ethan Stoner

GenAI Gateway on AWS — Production Deployment

AWS Terraform Docker

Deployed: Took AWS's multi-provider Generative AI gateway from an empty repo to a live, TLS-secured production endpoint — provisioning ~120 AWS resources with Terraform and owning the full deployment lifecycle. An OpenAI-compatible API layer (LiteLLM) sits in front of Amazon Bedrock, exposing Amazon Nova and Anthropic Claude through a single endpoint with usage-based routing and automatic fallbacks.

Hands-On Work: Provisioned VPC, ECS Fargate, ALB, PostgreSQL on RDS, Redis on ElastiCache, WAF, Secrets Manager, and Bedrock VPC endpoints via Terraform with a remote S3 state backend for reproducible, team-managed infra. Custom domain on Route 53 + DNS-validated ACM certificate; horizontal autoscaling, least-privilege IAM, Redis caching, and per-key cost/token observability. Layered Bedrock Guardrails + a Microsoft Presidio sidecar for PII masking — then traced a masking defect to a dependency version and fixed it via a controlled upgrade and regression testing. Built one-command spin-up / spin-down cost controls and rebuilt the entire stack from Terraform state to verify it.

Scope & credit: A real production cloud/DevOps task set by Cisco professionals at Riffyx Labs and completed end to end. Hands-on deployment and operation of AWS's official “Guidance for Multi-Provider Generative AI Gateway on AWS” reference architecture — infrastructure and ops work, not original authorship of the codebase.

~120 Resources via Terraform Guardrails + PII Masking Spin-Up/Down Cost Controls

Landed — Pricing Platform for Card Shops

Cloudflare Workers TypeScript Stripe

Built: A live subscription web product (landedcards.com) that tells a physical card shop what to put on the sticker. Paste a card, get a whole-dollar price for all five conditions — priced on landed cost (item + shipping), so buying it in the case beats waiting a week for the mail — with the evidence behind every number shown beside it. In production use by two shops.

Hands-On Work: Sole engineer across product, infrastructure, billing, and support. Cloudflare Workers + D1 + KV at the edge, Clerk auth with per-organization roles, Stripe subscription billing with plan entitlements and 14-day trials, and an hourly cron sweep that re-prices inventory and flags stickers that have drifted. The pricing engine is implemented twice on purpose — a Python reference and the TypeScript port that actually runs — with golden tests asserting the port reproduces the reference byte-for-byte from shared fixtures. 359 automated tests across both.

Also built: a companion intake tool that scans a stack of cards on a flatbed and identifies each one by artwork rather than by name — ORB keypoint matching, measured at 5/5 correct identifications with 4/4 correct refusals — then grades it from edge, corner, and centering defects measured in OpenCV and routes it to a human review queue before anything is listed.

Why It Matters: A real product with outside users, not a demo — auth, multi-tenancy, billing, rate limits, scheduled jobs, and the operational discipline that comes with a production database someone else's inventory lives in.

Live — Used by 2 Shops Subscription Billing & RBAC 359 Automated Tests

Pincer — Real-Time Computer-Vision Agent

Python PyTorch Computer Vision

Built: A real-time computer-vision agent that drives physical Android phones autonomously to play a mobile game — streaming live video off each device over USB, locating targets with a fine-tuned YOLO11 detector, and executing the on-screen gestures to act on them, closing the perception → decision → actuation loop end to end.

Hands-On Work: Fine-tuned YOLO11-small on 5,400+ self-collected frames via a self-supervised data flywheel — the running agent labels its own training data, corrected through a human-in-the-loop review UI, then retrained on a train → validation-gate → ship pipeline with ONNX export. Cut per-frame capture from ~600ms to ~1.3ms and each tap from ~200ms to ~20ms; multi-device orchestration with self-healing recovery and a live MJPEG monitoring dashboard.

Why It Matters: My most technically deep project — a complete perception-to-actuation loop with a self-improving ML pipeline, 127 automated tests, and structural safety invariants enforced in code and covered by tests, so it runs unattended for hours.

Real-Time Detection Self-Training Flywheel 127 Automated Tests Private Source

Qorlyt — Image-to-3D Generator

Python FastAPI WebSockets

Built: A self-hosted web product that turns a single image into a textured .glb 3D model in about 80 seconds via a local ComfyUI + Hunyuan3D 2.1 GPU pipeline.

Hands-On Work: A persistent WebSocket to ComfyUI maps per-node progress back to each user in real time; an async job queue serializes GPU work with live queue position and ETA; every generation is stored with an in-browser 3D viewer. Shared-password auth, GPU health pings, persistent history.

Why It Matters: End-to-end ownership of a real deployed AI product — web delivery, local inference, real-time systems, and reliability engineering.

GPU Inference Real-Time Progress 3D Generation

kvstore — Redis-Compatible Database

Java 21 RESP Protocol LSM-Tree

Built: A persistent, Redis-compatible key-value database written from scratch in Java 21 — an LSM-tree storage engine behind a TCP server that speaks the Redis RESP protocol, so standard Redis clients connect to it directly.

Hands-On Work: Built the LSM-tree engine end to end — write-ahead log, leveled compaction, bloom filters, and concurrent flush — plus TLS, multi-user auth, pub/sub, and snapshots. Covered by 203 automated tests.

Why It Matters: A systems-programming showcase — durable storage, concurrency control, and wire-protocol compatibility — the database and distributed-systems fundamentals that back real cloud infrastructure.

LSM-Tree Engine 203 Automated Tests RESP + TLS

Roblox UGC Pipeline

ComfyUI Live Revenue

Built: A self-hosted local-GPU ComfyUI pipeline that designs and renders Roblox UGC accessories end to end, then publishes them to the Roblox Marketplace.

Hands-On Work: Runs image-generation models entirely on local GPU hardware (no paid API), with a repeatable prompt-to-asset workflow for producing and publishing sellable items at volume.

Why It Matters: 11,000+ AI-generated models sold to date generating ~$150/month recurring — real, monetized proof of local model inference and GPU workload management at scale, not just a demo.

11K+ Units Sold Local GPU Inference Recurring Revenue

Lyric Generator

Python FastAPI ffmpeg

Built: Paste a Spotify link, get a brat-style lyric video. A FastAPI pipeline: Spotify metadata -> local / yt-dlp audio -> Whisper word-level timing (LRCLIB fallback) -> Pillow frame rendering -> ffmpeg mux.

Hands-On Work: Async job queue, graceful failure handling on every external call, packaged with pyproject.toml, MIT-licensed, and tested in GitHub Actions CI on every push.

Why It Matters: A polished, CI-tested, packaged Python project — production discipline end to end.

Whisper Timing CI Tested pyproject Packaged

yt2tiktok

Python ffmpeg / NVENC

Built: A desktop app that turns any YouTube video into scheduled, captioned vertical TikTok clips.

Hands-On Work: Splits into 60-70s segments (silence-based or LLM cliffhanger cuts), burns in word-by-word karaoke captions (6 style presets), NVENC GPU encoding with CPU fallback, parallel encoding (2-4x faster), and hybrid transcription — YouTube captions with automatic local Whisper fill-in for censored words.

Why It Matters: Real media-processing engineering: GPU acceleration, concurrency, and a polished desktop UI.

NVENC Encoding Karaoke Captions Parallel Encoding

CSUSM Campus Monitor

Python FastAPI OpenCV

Built: A real-time occupancy monitor for CSUSM campus locations that uses computer vision to detect people in live camera feeds and serve crowd data through a web dashboard.

Hands-On Work: Captured HLS video streams via ffmpeg, ran YOLOv8 person detection with custom false-positive filtering (rolling-window spatial tracker), built a FastAPI backend with SQLite time-series storage, and created an interactive dashboard with live video, heatmaps, and best-times recommendations. Dual detection backend supporting both local YOLO and Frigate NVR via MQTT.

Why It Matters: Solves a real campus problem and demonstrates computer vision, full-stack development, systems integration, and data engineering. Test suite with CI pipeline.

Computer Vision Real-Time Detection CI/CD

GSI Sand & Gravel

Next.js React

Built: A production marketing website for GSI Sand & Gravel, a Southern California sand & gravel supplier serving the region for 30+ years.

Hands-On Work: Responsive Next.js site presenting the company's services, coverage area, and contact funnel. Managed deployment, DNS, and ongoing maintenance.

Why It Matters: Real paid client work — shipped for an actual business.

Client Work Responsive Design DNS & Hosting

Hunyuan3D-2.1 Install Guide

13 Stars Docs

Built: A comprehensive, step-by-step guide for installing Hunyuan3D-2.1 with ComfyUI on Windows.

Hands-On Work: Documented the full GPU setup, dependency resolution, and model installation from real troubleshooting.

Why It Matters: My most-starred repo — used by the community to get a complex AI 3D pipeline running.

Technical Writing GPU Setup

rakazo — Windows Startup Hang

PR #322 Merged 1.4k Stars

The Bug: Three dev services silently failed to start on Windows under turbo — no output, no port, no error. The reporter could reproduce it but wrote "I cannot explain the root cause."

Hands-On Work: Narrowed it by elimination — a trivial script under turbo worked while the real app did not; redirecting the child's stdout to a file still hung, ruling out pipe blocking; a 160-second wait proved a hang rather than slowness. The cause was tsx watch's press-Return-to-restart stdin listener: under turbo on Windows that inherited handle never delivers and never closes, so the process never reaches its entrypoint.

Why It Matters: Fixed with a small cross-platform runner that detaches stdin while preserving output and signal semantics. Merged upstream.

Root-Cause Analysis Process I/O Node.js / TypeScript

TrueForge — Unbounded API Calls

PR #464 Merged 4.7k Stars

The Bug: A sandbox status refresh called a third-party API with no timeout from three request-handling paths, so a stalled provider could hold API requests open until the HTTP client's multi-minute default.

Hands-On Work: Added a timeout budget and an abort on the request itself. When a reviewer questioned whether the abort was necessary, I measured it rather than argued it — Promise.race frees the caller but leaves the socket open, demonstrated with socket-close events.

Why It Matters: Merged after a changes-requested review round in a 4.7k-star AI agent platform.

Timeouts / AbortSignal Code Review Node.js / TypeScript

Given a real production task by Cisco professionals at Riffyx Labs and completed it end to end — deployed and operated a production Generative AI gateway on AWS, provisioning ~120 resources with Terraform (ECS Fargate, RDS, ALB, WAF, Bedrock) with Guardrails and PII masking. (See the GenAI Gateway on AWS project above.)

Used Git-based workflows — branches, commits, and pull requests — plus hands-on production debugging to ship and maintain real infrastructure with experienced engineers.

AWS + Terraform Git Workflows Debugging

Design, build, and ship marketing websites for real paying clients — handling the full engagement from layout and build through deployment, DNS, and ongoing maintenance.

Most recent: a production Next.js site for GSI Sand & Gravel, a Southern California supplier serving the region for 30+ years.

Next.js Client Delivery DNS & Hosting

Represented the company at markets and events, answered customer questions, and handled cash and digital transactions accurately.

Also tested Stripe payment flows for the Cheryl's Caramels ecommerce site before release to verify checkout and payment reliability.

Customer Service Communication Stripe Testing

Co-founded a residential trash can power washing business and worked directly with customers from scheduling through service completion.

Managed scheduling, customer communication, cash handling, and income and expense tracking while delivering services reliably.

Entrepreneurship Customer Communication Scheduling