Automation Expert
16+ years building end-to-end automation systems across marketing, logistics, and e-commerce. Hands-on across the modern AI stack: LLM integration, RAG pipelines, agentic workflows, and voice AI. Full-stack and data engineering depth to ship them into production, not just prototype.
Macro-tracked OR calorie-tracked meal plans for Cebu City. Per-customer ingredient swaps within allergies and restrictions, group ordering with bulk discounts (gym buddies / officemates / condo clusters), bundled cold brew and matcha. Real business: pricing engine, schema, waitlist live; payments and kitchen ops layered in post-validation.
Browser voice chat. Web Speech API for in-browser STT and TTS, Groq behind a Vercel Edge function. Zero per-minute cost (no Twilio, Bland.ai, or ElevenLabs).
Vertical slice of an operational MVP for a Trucking OS. Driver list with computed scores and tiers, driver detail with score breakdown, and "post a load → LLM ranks the top 3 drivers with reasoning" flow. Built for a specific dispatch/fleet role; data layer structured to swap to Supabase.
A daily snapshot of Hacker News, classified by an LLM into categories with tech tags and a clickbait flag. The pipeline runs on a Vercel Cron Job that pulls top stories from HN's API, calls Groq for each title, and refreshes Next.js's Data Cache so the page is always pre-warm.
Scrapes the latest HN "Who is hiring?" thread, parses each top-level comment into structured fields with an LLM (company, role, remote status, comp, tech stack), and scores each posting 1-10 for fit against my resume. Daily Vercel cron, batched LLM calls (8 jobs per request) for throughput.
A GitHub Action that reviews pull requests with an LLM. Reads the PR diff, asks Groq for a focused review (bugs, security, edge cases, performance), and posts the result as a PR comment. Zero npm dependencies, vanilla Node 20. Used by other repos via uses: ramil-lina36/portfolio/pr-reviewer@main.
Retrieval-Augmented Generation over a marketing strategy corpus with inline numbered citations. Local ONNX embeddings, ChromaDB vector store, Groq generation, Gradio UI. The full RAG loop in ~150 lines, no LangChain.
Playwright scraper for the Hacker News front page with LLM-driven enrichment (category, tech tags, clickbait flag) into typed Pydantic models, persisted to SQLite. Pattern: deterministic scraping for what HTML gives you, LLM for what it doesn't.
The same lead-qualification workflow built two ways: as an n8n graph and as a Python agent with structured LLM output. A practical comparison of visual orchestrators vs code-based agentic systems.
Python, FastAPI, Next.js, TypeScript, Pydantic, Anthropic, OpenAI, Groq, ElevenLabs, Bland.ai, Playwright, ChromaDB, Supabase, PostgreSQL, BigQuery, n8n, Zapier, Make, Stripe, Shopify, HubSpot, EDI (X12, EDIFACT, SPS Commerce), Vercel, Hugging Face Spaces, Modal, Digital Ocean.