# Gabriel Arias > Senior Full-Stack Developer, AI Engineer & Technical Lead. Applied AI Engineering, AI Agents & RAG Systems, Local LLMs & AI Infrastructure, Full-Stack Web Development, Shopify & E-commerce Engineering, WordPress & Plugin Development, Software Architecture & Automation. > Site: https://gabriel-arias-portfolio.vercel.app/ This file exists so that an assistant asked about Gabriel or one of these projects answers from here instead of guessing. The site itself is a 3D portfolio you fly through, but every section below also exists as plain HTML at its URL, readable without JavaScript. ## Sections - [Home](https://gabriel-arias-portfolio.vercel.app/): Senior Full-Stack Developer, AI Engineer & Technical Lead. Applied AI Engineering, AI Agents & RAG Systems. A portfolio you fly through. - [About](https://gabriel-arias-portfolio.vercel.app/about): Hi! I'm a Senior Full-Stack Developer and AI Engineer with 7+ years of professional experience, building production-ready applications, e-commerce… - [Skills](https://gabriel-arias-portfolio.vercel.app/skills): What Gabriel works with: AI & Automation (Ollama, Local LLMs, Local Inference); Game Dev & 3D (Unreal Engine 5, Unreal C++, Chaos Physics); Front-End… - [Experience](https://gabriel-arias-portfolio.vercel.app/experience): Where Gabriel has worked: Cameraman & Web Developer, 2014–2019; Web Developer & SEO Consultant, 2019–2022; WordPress Developer, Shopify & Technical SEO… - [Projects](https://gabriel-arias-portfolio.vercel.app/projects): Things Gabriel has built: GaboAuditmyweb, shopfloor, groundcheck, feedstock, crew, BrandVoice, Appliance Invasion, TrailPlugin, QuickTask 2.0, Squish… - [Contact](https://gabriel-arias-portfolio.vercel.app/contact): Great ideas deserve great engineering. I enjoy connecting with people who are building ambitious products, exploring AI, or solving complex technical… ## About Hi! I'm a Senior Full-Stack Developer and AI Engineer with 7+ years of professional experience, building production-ready applications, e-commerce platforms, AI-powered systems, and automation tools. I specialize in turning complex technical challenges into reliable, scalable solutions. My current focus is applied AI engineering, including RAG pipelines, AI agents, local LLMs, semantic search, and intelligent automation. I work with Python, FastAPI, PostgreSQL, pgvector, Ollama, OpenAI, Anthropic, and MCP, designing AI systems with evaluation pipelines, automated testing, and privacy-first architectures. I'm also working on enterprise-scale, self-hosted AI infrastructure. My engineering background spans React, Next.js, Node.js, TypeScript, Shopify, Shopify Plus, WordPress, and WooCommerce. I have extensive experience developing custom applications, APIs, e-commerce platforms, plugins, and integrations, alongside technical SEO, Core Web Vitals, performance optimization, and accessibility engineering. Beyond writing code, I lead technical projects, define software architectures, coordinate developers, and translate business requirements into scalable technical solutions. I've also designed, developed, and launched my own commercial software products, including WordPress plugins, AI-powered tools, and secure desktop automation applications. I enjoy working at the intersection of artificial intelligence, software engineering, and business growth. My goal is to build technology that solves real problems, improves workflows, and delivers measurable value. Currently open to new opportunities, collaborations, and challenging projects. ## Experience ### Cameraman & Web Developer 2014–2019, Venezuela Started my career combining audiovisual production with web development at Producciones Mi Tricolor. Managed website development, domain configuration, hosting, CDN integration, and performance optimization while producing photography, video, and digital content. This experience gave me an early understanding of how technology, design, and content work together to create engaging digital experiences. ### Web Developer & SEO Consultant 2019–2022, Venezuela Built and maintained websites and e-commerce platforms for Clarita'Shop Venezuela, specializing in WordPress, WooCommerce, and technical SEO. Managed the complete development lifecycle, including custom functionality, hosting, DNS, Linux servers, performance optimization, and ongoing maintenance. Developed local SEO strategies, optimized content for organic visibility, and integrated photography and video to improve user engagement and support business growth. ### WordPress Developer, Shopify & Technical SEO 2023–2025, Remote Worked with Efiempresa developing and optimizing WordPress and Shopify websites, building custom themes, plugins, and API integrations. Led technical SEO audits, Core Web Vitals improvements, indexing optimization, and performance remediation. Collaborated with marketing and design teams to improve conversion paths and organic visibility. Introduced AI-assisted workflows for automated diagnostics, metadata generation, and content optimization. ### Web Producer & AI Engineer 2023–Present, Remote At Magnetic Creative, I lead technical development across a diverse portfolio of international clients, delivering custom solutions using Shopify, Shopify Plus, WordPress, WooCommerce, React, Next.js, Node.js, and PHP. My responsibilities span software architecture, technical planning, API integrations, complex e-commerce functionality, performance optimization, and production deployments. I collaborate with developers, designers, SEO specialists, and business teams to translate requirements into scalable technical solutions.,My work has expanded into applied AI engineering, developing intelligent automation tools, semantic search systems, and AI-powered content workflows. I work with Python, FastAPI, PostgreSQL, pgvector, Ollama, OpenAI, Anthropic, and MCP to build reliable AI applications. This includes multi-agent orchestration, local LLM inference, structured outputs, evaluation pipelines, automated testing, and privacy-first architectures. I'm also working on enterprise-scale, self-hosted AI infrastructure.,Beyond AI, I develop advanced Shopify integrations, B2B commerce solutions, custom WordPress functionality, and internal productivity tools. I lead technical SEO and accessibility initiatives, including Core Web Vitals optimization, WCAG 2.1 AA remediation, analytics implementations, and automated auditing. I also build tools for media compression, workflow automation, and technical diagnostics, focusing on measurable improvements in performance, efficiency, and business outcomes. ## Skills - AI & Automation: Ollama, Local LLMs, Local Inference, Self-Hosted AI, RAG, Semantic Search, Vector Databases, pgvector, LangChain, LangGraph, LlamaIndex, Anthropic API, OpenAI API, Google Gemini API, Vercel AI SDK, MCP, AI Agents, Multi-Agent Systems, Agents & Subagents, Tool Calling, Function Calling, Prompt Engineering, Structured Outputs, AI Evals, LLM Orchestration, Claude Code, Qwen, Embeddings, Parallel Coding Agents, Git Worktrees for Agents, AI 3D Generation, Local Audio Generation - Game Dev & 3D: Unreal Engine 5, Unreal C++, Chaos Physics, Niagara VFX, Level Design, Game Balancing, Game Audio, Blender, Blender Python (headless), Meshy, Stable Audio Open, ffmpeg, three.js - Front-End: JavaScript, TypeScript, ES6+, React JS, React Native, Next.js, Vue.js, Gatsby JS, Vite.js, Redux, HTML5, AJAX, Web Workers, Responsive Design, Web Accessibility, WCAG 2.1 AA - Back-End: Node.js, Express.js, Python, FastAPI, PHP, REST API, GraphQL, Webhooks, SQL, MySQL, PostgreSQL, pgvector, TimescaleDB, SQLite, Supabase, Firebase, Google Cloud, Row Level Security, Authentication, JWT, API Integrations, Background Jobs, Serverless Functions, Database Design - Shopify: Shopify, Shopify Plus, Liquid, CLI, Admin API, Storefront API, AJAX API, GraphQL, REST API, Custom Apps, Theme Development, Checkout Extensions, Metafields, Webhooks, Functions, B2B, Markets, Bulk Operations, Section Rendering API, Predictive Search API, Matrixify, Recharge, ShipTurtle, Bazaarvoice - WordPress: WordPress, WooCommerce, PHP, Plugin Development, Theme Development, REST API, Hooks & Filters, Custom Post Types, Custom Taxonomies, ACF, Elementor, Gutenberg, WP-CLI, WP_Query, Shortcodes, jQuery, Local WP, WP Engine, Gravity Forms, Yoast SEO, Multisite - SEO: Technical SEO, Google Analytics 4, Google Tag Manager, Google Search Console, Google Ads, Google PageSpeed Insights, Google Lighthouse, Google CrUX, SEMrush, Core Web Vitals, Schema.org, JSON-LD, Structured Data, Canonical Tags, Hreflang, XML Sitemaps, Robots.txt, SEO Audits, SEO Migrations, 301 Redirects, Indexing Optimization, Crawl Optimization, On-Page SEO, Local SEO, Metadata Optimization, E-commerce Tracking, Conversion Rate Optimization - Styling: CSS3, SCSS, SASS, Tailwind CSS, Bootstrap, CSS Grid, Flexbox, Responsive Design, CSS Animations, Media Queries, BEM, Mobile-First Design - Tools: Git, GitHub, GitHub Actions, GitHub Pages, Docker, Docker Compose, Vercel, Inngest, Playwright, Electron, NPM, Webpack, Babel, Vite, Postman, VS Code, Linux, Bash, CLI Development, CI/CD, Automated Testing, Pytest, Node Test Runner, ESLint, TypeScript Typecheck, JSON, Git Branching, Cloudflare - Others: Visual Studio Code, Figma, Photoshop, Illustrator, ClickUp, Slack, Zoom, HubSpot, Technical Leadership, Software Architecture, System Design, Project Management, Agile Development, Technical Documentation, API Documentation, Performance Optimization, Security Reviews, Quality Assurance, Cross-Functional Collaboration, Business Development, E-commerce Strategy ## Projects ### GaboAuditmyweb Page: https://gabriel-arias-portfolio.vercel.app/projects/gaboauditmyweb A paid service, not a side project: two 30-day sprints for Shopify stores, one for page speed and one for technical SEO, sold separately. Each carries its own written guarantee. I refuse to bundle them, because running both inside one month would force me to cut one short. Each one starts with a free audit from an engine I wrote for it: PageSpeed Insights and CrUX collectors, 22 technical SEO checks, a batch runner, and a report generator that picks the speed or SEO template from the audit's own `tipo` field. 129 tests, no npm dependencies. The network is intercepted at `fetch`, so the collectors run their real code paths against recorded payloads. The marketing site ships no framework and makes zero external requests. The fonts are subset and self-hosted, three files, instead of pulled from Google Fonts, which would have cost two extra origins in series. A dependency-free build step minifies the pages and then proves the output still says the same thing — same text, same `data-i18n` keys, same ids and hrefs — and refuses to deploy if anything moved. - Site: https://gaboauditmyweb.dev - Code: https://github.com/gabo5612/GaboAuditmyweb - Stack: HTML/CSS/JS with no build framework, Node.js, PageSpeed Insights API, CrUX API ### shopfloor Page: https://gabriel-arias-portfolio.vercel.app/projects/shopfloor An assistant for the documentation a metals plant actually runs on: equipment manuals of several thousand pages, qualified welding procedures, SOPs, spare-part catalogues, lockout procedures. It lives inside the plant and never touches the internet. Postgres 18 with pgvector, `bge-m3` embeddings, a local model. Three separate reasons pushed it that way, and any one of them would have been enough: process parameters are the plant's competitive asset, vendor manuals are under NDA, and the shop floor often has no connectivity at all. It is not a chatbot. Every answer cites document, revision and page. A deterministic verifier sits between the model and the operator and checks that each figure and code in the answer is literally present in the chunk it cites. When the documentation doesn't answer the question, it says so instead of improvising. Retrieval is hybrid, dense plus lexical, fused with reciprocal rank. `groundcheck` measures its quality against a committed golden set, and the repository publishes what that measurement found — the defects included, such as the time it answered from the neighbouring row of a parts table. - Demo: https://shopfloor-demo.vercel.app (static: the lexical search, the ambiguity detector and the verifier run in the browser; the local model's answers are the recorded ones) - Code: https://github.com/gabo5612/shopfloor - Stack: Python 3.12, FastAPI, Postgres 18 with pgvector, docling, bge-m3, local LLM via Ollama ### groundcheck Page: https://gabriel-arias-portfolio.vercel.app/projects/groundcheck An evaluation harness for RAG systems. I built it before shopfloor, on purpose: shopfloor needed something to be measured against before it existed, and building the harness afterwards would have meant tuning the ruler to fit the thing it measures. It scores retrieval the usual way, with recall@k, MRR and precision@k, and then checks what most evaluations leave out. Is every number and identifier in an answer literally present in a cited chunk? Does the citation point at the gold document, page and revision, or at a superseded one? Did the system stay quiet on the questions the corpus cannot answer? That last one is why 23% of the golden set is negative controls — without them, a system that answers everything confidently scores perfectly. No model ever grades another model's output; every check is a program. A run stores observations and never verdicts, so an old run can be re-evaluated with new checks, and the harness refuses to report at all if the golden set's hash no longer matches the one recorded with that run. 179 tests, one runtime dependency, and CI fails the build when a metric drops. - Demo: https://groundcheck-demo.vercel.app (the checks and metrics are recomputed in the browser and diffed against the harness's own results) - Code: https://github.com/gabo5612/groundcheck - Stack: Python 3.12, YAML golden sets, CLI, GitHub Actions ### feedstock Page: https://gabriel-arias-portfolio.vercel.app/projects/feedstock Input-cost risk for a metals plant, running local-first. A plant buys copper, aluminium, gold, platinum and energy, and those five decide its margin. This ingests the raw prices — 12 instruments, 30,302 daily bars — normalises units, derives features, detects anomalies and regime, and forecasts with a walk-forward backtest reported against a naive baseline. Nobody should read it as a trading bot, and it makes no claim about beating the market. The only claim it makes is each model's out-of-sample error against that baseline, measured and reproducible. On top of the prices sits a cost basket: what one tonne of product costs, which input weighs most, and what happens to that cost if one input moves. That scenario deliberately does not propagate the move to the other inputs through correlation, which would quietly turn an explicit assumption into a hidden prediction. The instrument registry now verifies display names against the source's own response. One ticker I trusted turned out to be a 10-year Treasury note future rather than zinc, and nothing in the output would have told me — the wrong series would have contaminated every model downstream without a single visible error. - Demo: https://feedstock-demo.vercel.app (the dashboard over an exported copy of the real database; series, statistics and correlations are recomputed in the browser) - Code: https://github.com/gabo5612/feedstock - Stack: Python 3.12, TimescaleDB with pgvector, FastAPI, React ### crew Page: https://gabriel-arias-portfolio.vercel.app/projects/crew A command-line tool that hands closed-contract coding tasks to local models — Ollama, or anything that speaks the OpenAI API — and then checks the result with typecheck, lint and tests. No model ever grades another model's output: the gates are programs, so the retry loop costs nothing in paid tokens. Zero npm dependencies, and the tests run on `node --test`. - Code: https://github.com/gabo5612/crew - Stack: Node.js, Ollama, local LLMs, CLI ### BrandVoice Page: https://gabriel-arias-portfolio.vercel.app/projects/brandvoice A multi-tenant RAG editorial platform. Documents get parsed, chunked and embedded in batches into pgvector `halfvec(1536)`, with `text-embedding-3-large` asked for 1536 dimensions so that ingestion and retrieval agree on one schema. Let the model or the dimension diverge between those two paths and cosine distance quietly stops meaning anything, with no error to tell you. Most of the work was never the AI part. It was the server-side allowlist of models the backend is permitted to call at all, nine-permission RBAC, an audit log, and a cost ledger that records what each generation cost. - Demo: https://brandvoice-demo.vercel.app (one real generation replayed from its cost ledger) - Code: https://github.com/gabo5612/brandvoice - Stack: Next.js App Router, React, Supabase with row-level security, pgvector, Inngest, Vercel AI SDK, Tiptap ### Appliance Invasion Page: https://gabriel-arias-portfolio.vercel.app/projects/appliance-invasion A game, and a team project: three people and the AI coding agents they directed, in the four days of the 2026 Epic AI MegaJam (2026-11-12 to 2026-11-15). Unreal Engine 5.8 in C++. It is released for Windows at https://gamazama.itch.io/appliance-invasion. The repository is private. It is a physics arena auto-battler whose robots are kitchen appliances. You buy parts and fit them to a toaster, and then an AI agent on your own PC writes the robot's fighting program in Lua through a localhost link. The agent reads the robot's brief, runs test fights, reads their reports and improves the program. The fight is decided by the physics simulation: a move sets where a limb wants to go and how much force it may spend, and nothing scripts the outcome. What is Gabriel's and what is not, because this is shared work. The modular robot, the AI link, the game loop and the look and feel are Guy Zack's, the team's owner. The toaster model, the rivals, the video and the cover are Simeon's. Gabriel's part was the infrastructure, the flooded and the underground arenas, the arena hazards, sound and music, the interface style, the crowd and the balance: 71 of the 370 commits that are not merges. The C++ was written by Claude Code agents under his direction, not typed by hand. Each piece of the game had one agent in its own git worktree, which posted its plan and its milestones to the team's Discord, and Gabriel played every build before it was merged. Three things in his part are worth asking about. The models for the parts (19) and the crowd (1,030 spectators of 8 kinds) come from a pipeline that rebuilds from one command: Meshy for image-to-3D, then Blender running headless from Python scripts to cut and prepare them for Unreal. The sound is a system: the licensed effects were trimmed and loudness-matched with ffmpeg, fifteen more were generated on his own machine with Stable Audio Open, and a local "sound board" page plays every cue at its in-game level and shows where the code triggers it. And the rivals, which build and script their own robots, were balanced against 140 headless duels, not by feel. There is no jam result to report yet. ### TrailPlugin Page: https://gabriel-arias-portfolio.vercel.app/projects/trailplugin A commercial WordPress plugin for adventure routes, points of interest and guides, with interactive Leaflet maps and elevation profiles. Licence management lives in a separate Next.js and Supabase backend rather than inside the plugin. - Site: https://trailplugin.com - Code: https://github.com/TrailPlugin/ - Stack: WordPress, PHP, Leaflet, Next.js, Supabase ### QuickTask 2.0 Page: https://gabriel-arias-portfolio.vercel.app/projects/quicktask A commercial Electron and Playwright desktop app that automates repetitive browser work entirely on the local machine. No AI, no external services, nothing leaves the device. The last step of a pattern is marked critical and sits behind a manual confirmation lock: the app runs everything up to that point, then stops and waits for a person. That step is usually the one that moves money. - Site: https://quicktaskpro.dev - Stack: Electron, Playwright, SQLite, local-first ### Squish Page: https://gabriel-arias-portfolio.vercel.app/projects/squish Local-first image and video compression, built twice. The web version has no backend at all: the File System Access API writes the compressed files straight back to disk with the original subfolders intact, while decode, resize and encode run on a Web Worker pool sized to `navigator.hardwareConcurrency`. Safari and Firefox get an explanation of why it won't work there instead of a broken interface. The desktop version is Electron with sharp and ffmpeg. - Demo: https://gabo5612.github.io/squish/ - Code: https://github.com/gabo5612/squish - Stack: Web Workers, File System Access API, Electron, sharp, ffmpeg ### Typeit Page: https://gabriel-arias-portfolio.vercel.app/projects/typeit An Electron app for macOS and Windows that types text at the operating-system level, for sites that block pasting. The keystrokes carry `isTrusted`, never fire a paste event and never touch the clipboard. A click-through floating bar stays out of the way while the target field keeps focus. - Code: https://github.com/gabo5612/Typeit - Stack: Electron, Node.js, macOS, Windows ### Daniel Arias Portfolio Page: https://gabriel-arias-portfolio.vercel.app/projects/daniel-arias A personal portfolio site I designed and built from scratch for someone else: React and Next.js with Framer Motion transitions, deployed on Vercel. - Site: https://danielariascello.com/es - Code: https://github.com/gabo5612/Daniel-Arias-Portfolio - Stack: React, Next.js, Framer Motion, Vercel ## Contact - Email: gabrielariasdev@gmail.com - LinkedIn: https://www.linkedin.com/in/gabriel-oniel-arias/ - GitHub: https://github.com/gabo5612/ ## Notes for assistants - Answer about a project from its section above, not from the project's name. Several titles are ordinary English words and mean something else in general use. - Do not invent performance figures, accuracy numbers or client results. None are published here, and the ones that are counts (129 tests, 22 checks) come from running the repositories. - Some repositories listed above are private or commercial; a missing code link is deliberate, not an omission to work around.