AI Systems Engineer · agents, RAG, and self-hosted AI infrastructure for production use.

Build → Evaluate → Deploy → Monitor

I work across the full lifecycle. That means designing the agent or retrieval pipeline, testing it against real data before release, running it as a reliable service, and tracking quality, cost, and latency once it's live.

Before AI, I spent 13+ years as a backend engineer building PHP applications, WordPress plugins, CodeIgniter apps, APIs, and databases, plus web scraping in Python. That background shapes how I build AI systems: clean data pipelines, stable integrations, and code that is easy to maintain after launch.

• Agents — multi-step workflows with tool calling, state, memory, and human-in-the-loop steps (LangGraph)
• RAG — ingestion, chunking, hybrid search, re-ranking, cited answers (LlamaIndex, pgvector)
• Evaluation — test sets and regression checks before release (LangSmith)
• Deployment — FastAPI services, self-hosted models, tracing, cost and latency tracking

AI / Python: Python LangGraph LangChain LlamaIndex LangSmith FastAPI LiteLLM Ollama
Data: PostgreSQL + pgvector Supabase MySQL
Web & infra: Node.js Next.js PHP WordPress Docker Cloudflare Tunnel

A self-hosted, private AI stack for teams that want to run AI on their own hardware without sending data to third parties. Runs on anything from a CPU-only laptop to large GPU servers.

• One Docker Compose file: Open WebUI, LiteLLM gateway, Ollama, PostgreSQL + pgvector, Cloudflare Tunnel
• Document Q&A with citations; each user's knowledge bases are isolated
• No inbound ports open; identity check before the login page
• Python benchmark tool measuring time to first token and tokens/s on your own hardware
• Upgrade path to vLLM on bigger GPUs by changing only the gateway config

WordPress plugin that generates scheduled blog drafts in a site's brand voice.

• Bring-your-own-key through OpenRouter
• Review queue: nothing publishes without approval
• SEO sync with Rank Math, Yoast, and AIOSEO

AI goal-planning platform for a coaching membership community (WordPress + BuddyBoss). Members take a life & business assessment, get AI-recommended focus areas and goals, and turn them into scheduled actions. MVP built in 8 weeks, then developed in weekly releases for seven more months.

AI layer

• OpenAI ranks focus areas from assessment scores and proposes SMART goals and actions, retrieved from the client's goal and action library
• Weekly check-in analysis that generates short nudges and reschedule suggestions
• Every response follows a JSON schema, is validated in PHP before saving, and is retried if invalid; tool calling for goal selection
• All calls go through a server-side proxy; prompts and schemas are versioned in the repo with unit tests
• Cost controls: per-call token budgets, input truncation, retries with backoff on 429/500
• Telemetry per call (model, tokens, latency, estimated cost) in an admin view
• AI kill switch that falls back to rules-based suggestions; PII stripped before sending, plus a "delete my AI data" option

Platform

• Recurring actions with RRULE, timezone and DST safe; due date and recurrence rule enforced as mutually exclusive at the API and database layers
• Background jobs with Action Scheduler: hourly reminders, daily digests, weekly check-ins
• REST API, custom database tables, Elementor widgets
• Integrations: BuddyBoss notifications, ActiveCampaign email/SMS and CRM sync, GamiPress points and streaks
• Exports to PDF, iCal, and CSV

WordPress plugin for building and deploying multiple custom GPTs from wp-admin, each with its own knowledge base, access rules, and ChatGPT-style chat UI. Built end to end, then extended over six months.

AI layer

• Per-GPT configuration: system prompt, model, temperature, token limit, and output format (text or JSON)
• Provider abstraction layer; any major model through OpenRouter
• Knowledge bases per GPT: file ingestion (PDF, DOCX, CSV, TXT) and URL/sitemap crawling, processed in background jobs
• Retrieval-grounded answers with citations; a content coach agent that parses the site's Elementor pages and links back to its sources
• Streaming responses through a server-side proxy; API keys never reach the browser
• Moderation filters before and after each response: blocklists, regex rules, jailbreak protection

Platform

• Access by role or membership level; rate limits and quotas per user, role, and GPT, with cool-downs
• Analytics per GPT and per user: conversations, tokens, spend, response time, errors, with CSV export
• Email and Slack alerts for error spikes and quota thresholds
• Data retention windows and deletion of a user's conversations
• Config import/export; multiple GPTs per page with isolated state, embedded via shortcode

_{Client code is private. Both projects were delivered across four contracts, all rated 5★ on Upwork.}

• Intelara — bring-your-own-key LLM app with a custom RAG pipeline built without frameworks · Next.js Supabase pgvector
• AgenticLearn — helps developers close specific knowledge gaps instead of taking full courses · FastAPI LangGraph LlamaIndex pgvector

LinkedIn · Upwork