A decade of infrastructure and reliability engineering, and the AI systems I build now - sorted into folders. Open any one of them.
Open the drawer↓
The files · 2017 → present
Pick a folder
Systems that remove repeated human work: answering from your own documents, moving data between tools, and running multi-step jobs without a person in the middle. Built API-first, instrumented from day one.
Knowledge systems (RAG)
Problem
Staff re-answer the same questions from documents nobody can find.
Build
Question-answering over your own SOPs and support history. Your team edits the source document and the answers update immediately - no developer, no redeploy.
Scale
Direct context with prompt caching for small corpora; pgvector with change-detection re-indexing when the document count outgrows it.
Claude API · MCP · pgvector · embeddings & chunking
Process automation
Problem
Copying data between systems, routing tickets, rebuilding the same report every Monday.
Build
Workflows connecting the tools you already pay for, with the manual step taken out of the middle.
Measure
Hours returned per week - not how clever the workflow is.
n8n · Zapier · Make · REST & webhooks · scheduled jobs
Custom agent tooling
Problem
General-purpose assistants lose context between sessions and repeat mistakes on long-running work.
Build
Custom agent skills and MCP connectors giving the model durable access to project state, documented procedures and external tools.
Result
Multi-step work that survives between sessions instead of restarting from zero.
Claude API · Model Context Protocol · Agent Skills · in daily use
Ten years of finding out why production broke - hypervisor, storage, network, application. The same discipline now applied to AI systems, where the failures are silent and your monitoring stays green.
Root cause analysis on production incidents. Built internal tooling for field deployment and auto-support. Ran proactive risk analysis across the installed base to catch at-risk deployments before customers felt them.
An LLM giving a confident wrong answer returns HTTP 200 and logs nothing. You find out when a customer complains.
Build
Instrumentation the way infrastructure gets instrumented: request tracing, cost per call, latency percentiles, and eval suites that catch quality regressions before users do.
Why me
Most people building with LLMs have never had to keep a system running at 3am. I have.
Two live sites for small businesses, both hand-built and both in production. No CMS to maintain, no monthly platform bill, no page builder - build, deploy, DNS and forms owned end to end.
Temazcal - Casa de Vapor Zumpango
Client
Traditional temazcal and wellness practice in Zumpango, Estado de México.
Build
Three-section site in purple, black and white with a full-bleed video hero and scroll parallax. Services cover temazcal, masaje, ayahuasca, pariká and kambó; a monthly events page carries the flyer; contact routes straight to WhatsApp.
Stack
Hand-written HTML, CSS and JavaScript. No framework, no dependencies.
Cosmetic and general dental practice in Acapulco - smile design, orthodontics, whitening, aesthetic resins.
Build
Two-page site built around the practice's own clinical photography and video, with a services breakdown per treatment. Positioning line: every smile needs a different plan.
Stack
Hand-written HTML and CSS, optimized media, Instagram as the social channel.
Production tooling I built for my own channels - the parts of video work that are repetitive enough to be code.
Narration-locked video assembly
Problem
Scene timing is normally eyeballed, so visuals and narration drift apart over a thirteen-minute runtime.
Build
Word-level forced alignment derives every scene duration from the voiceover itself. Ken Burns on stills, xfade chaining with drift-free offset math, music ducked to −14 LUFS, H.264 render, scripted QA pass.
Result
Timing became data instead of judgment. Full-length videos assemble unattended and land in sync.
Python · faster-whisper · FFmpeg · automated QA
Ranked countdown generator
Problem
Building a top-ten countdown by hand means collecting clips, ordering them, and rebuilding the same template every time.
Build
A local service that ingests source links, normalizes clips, and renders the finished vertical video programmatically - React components compiled into video frames instead of a timeline.
Result
A list of links goes in, a finished Short comes out.
The engineering degree, the certifications, and seven years of support and operations before the reliability work - enterprise customers, real deadlines, and the habit of owning a problem until it closes.
Technical Support Specialist · Omnitracs
Scope
L1/L2 support across fleet telematics hardware and software for enterprise customers.
Work
Owned cases end to end through resolution, configured remote access solutions, led IT onboarding for new clients.
Span
Feb 2020 - Apr 2022 · 27 months
Tech Support Analyst · Tech Mahindra
Scope
US enterprise clients under strict real-time resolution targets.
Work
Mobile device fleets and the Target fulfillment platform, plus data analysis and reporting on support operations.
Span
Mar 2019 - Jan 2020 · 11 months
B.S. Computer Engineering
Degree
Bachelor of Science in Computer Engineering - Universidad Lucerna. The formal grounding under everything in these folders: systems, networks, databases and programming.
Certs
Nutanix Certified Master - Multicloud Infrastructure (NCM-MCI). Microsoft Technology Associate in Network Security Fundamentals, SQL Database Administration and Web Design Fundamentals. Scrum and Project Management.
Languages
English and Spanish, both native - raised and educated in the United States.
Budget administration and financial tracking for logistics IT, including warehouse systems for Ford.
Work
Supervised the dispatch team, first point of contact for client requests, and built Excel and Visual Basic automation for inventory and earnings reporting.
Span
Mar 2017 - Dec 2017 · 10 months
Contact
So what can we work on?
Or what's taking your team ten hours a week that shouldn't. I reply to everything, usually same day.