Work Experience

Recent AI Work

• Claude-Powered Enrichment & Qualification Agent: End-to-end agent built on the Anthropic API + Agent SDK that takes a registrant record, runs web/LinkedIn verification through custom MCP tools, applies a versioned rubric, and emits a qualification decision plus reviewer-ready summary — replacing ~20 minutes of manual operator work per record.
• Custom MCP Connectors: Built MCP servers (Python and TypeScript) wrapping platforms without official servers — typed tool definitions, idempotent operations, OAuth/secrets handling, and rate-limit-aware retries. Pattern is reusable for Luma, Airtable, GHL, and arbitrary REST/GraphQL APIs.
• Reviewer Surfaces on Vercel: Lightweight Next.js apps deployed on Vercel for one-click human approval, with magic-link auth, Airtable-backed audit logs, and Slack notifications. Picks the lightest surface (email handler → Airtable Interface → Next.js app) that solves the use case.
• Agent Eval Harnesses: Versioned prompts and rubrics with regression test sets, Langfuse / LangSmith tracing, and CI checks that block deploys when eval scores drop — quality as a testable property.
• Workflow AI Upgrades: Audited Make / Zapier / GHL / Airtable flows, replaced brittle conditional logic with Claude-driven classification and personalization where it raised the ceiling; left routine automations untouched.
• Claude Code in Daily Workflow: Uses Claude Code to build, test, and maintain production systems — including scaffolding test harnesses, generating MCP boilerplate, and writing runbooks the Automations Specialist can take over from.
• Idempotent Webhook Pipelines: Webhook receivers (FastAPI / Node.js) with deduplication keys, exponential-backoff retries, dead-letter queues, and structured logging routed to error tracking — no half-applied state on partial failures.
• SAM.gov Procurement Pipelines: Continuous ingest, classify, and alert; vector-store routing across PGVector/Pinecone/Chroma with cost/latency-aware policies.

Claude Code & Cowork — Daily Practice

Uses Anthropic's full developer surface every day — not just the API. Builds, ships, and operates production AI systems through Claude Code (CLI), the Agent SDK, the Anthropic API, and Cowork (the desktop agent product). What you see on this CV — the PHP edits, the bilingual translation arrays, the knowledge matrix — is itself produced and maintained through a Cowork session driving Claude against a real workspace folder, with file tools, a sandboxed Linux shell, and connector MCPs.

• Claude Code as primary engineering tool: scaffolds, refactors, tests, and maintains production code across Python, TypeScript, PHP, and C# repositories. Daily driver — not a chat companion. Comfortable with sub-agents (Task / Agent tool), slash commands, hooks, project- and user-level settings, and MCP servers configured per project.
• Custom Cowork plugins: builds and customizes Cowork plugins (bundles of skills, MCP servers, slash commands, and sub-agents) tailored to specific operator workflows. Comfortable scaffolding a plugin from scratch and customizing existing ones (HR, design, infographic generation, plugin authoring itself).
• Skill authoring (SKILL.md): writes Claude skills using the progressive-disclosure pattern — frontmatter triggers tuned for accuracy, a short SKILL.md core, bundled scripts loaded only when relevant, and eval cases that verify the skill triggers on the right prompts and not on look-alikes.
• Agent SDK + custom MCP: spins up agents with file tools, sandboxed shell access, and custom MCP tools — the same control surface Claude Code exposes. Builds MCP servers in Python and TypeScript that wrap arbitrary REST/GraphQL APIs as typed tools (e.g. Luma, Airtable, GHL patterns) when no official server exists.
• Computer use & Claude in Chrome: drives native desktop apps (Finder, Mail, Notes, third-party apps) and web apps from agents (computer-use MCP, Chrome extension) when the workflow crosses apps or has no API. Knows when to drop from a dedicated MCP → Chrome MCP → computer-use, and when not to.
• Live artifacts: ships persisted HTML artifacts backed by connector data — refresh-on-open, in-page Haiku inference for summarization/classification, and scheduled-task triggers — for status pages, project trackers, and recurring digests the user re-opens daily.
• Scheduled tasks: automates recurring work through Cowork scheduling (daily briefings, recurring reports, "remind me in an hour", "run this at 6am"). Knows the difference between a one-off and a cadence — and offers to schedule proactively when the request is naturally recurring.
• MCP connector ecosystem: uses connectors across Slack, Notion, Atlassian, MS365, Figma, Linear, Intercom, Asana, Google Workspace, and more. Searches the MCP registry before falling back to browser automation. Builds custom servers when no connector exists.
• AskUserQuestion + TaskList discipline: clarifies underspecified requests up front, tracks multi-step work in a structured task list, and runs verification steps (re-reads, programmatic checks, second-opinion sub-agents) before declaring "done."

Selected Projects

• Multiple Protein Sequence Alignment (C#, CUDA, C++): Guide-tree innovation using novel radial distance; repositories: PhdThesis • Align4
• Icarus Softcore (VHDL, VB, Java): Softcore CPU for FPGAs; compiler emits tailored hardware to reduce device utilization.
• Telegram Bots (PHP, MySQL, Web APIs): @TradingAlert_bot (Ichimoku signals), @AdsVideoPlayer_bot (remote playlist control), @ridesmx_bot (rides), @bookServerbot (67k+ titles; tagging/resume-based search).
• GymOnline (PHP, MySQL, Embedded): POS & access control across 5 gyms; Arduino bridges PC↔turnstiles; later WinForms app w/ fingerprint & printing.
• Basketball Scoreboard (VB.NET): Live scoring with dual display — demo: YouTube.
• Computer Vision (OpenCV, YOLO): RealSense-based pedestrian/vehicle detection; object counting on public transport.
• Ticket Vending Machines (TVM): Full stack incl. ChipDNA/Ingenico, QR tickets, coin/bill validators, RFID, printers; plus central control manager (C#).
• Semantic Code Intelligence Platform (Python, Next.js, PostgreSQL, Docker): Self-hosted platform that indexes multiple codebases using LLM-powered embeddings and AST parsing; features hybrid search (vector, full-text, trigram, BM25), automatic incremental indexing, and a web dashboard for code exploration.
• Smart Waste Collection Platform (Python, Vue.js, .NET MAUI, PostgreSQL/PostGIS, Redis, Docker): Multi-tenant municipal waste management system with real-time GPS tracking, route optimization (10+ algorithms including TSP, VRP, ACO), citizen engagement PWA, mobile collector and citizen apps, WebSocket live updates, and Celery background processing.
• Auction Intelligence Platform (Python, Next.js, PostgreSQL/pgvector, MinIO, Redis, Docker): Automated vehicle auction monitoring system with web scraping, asset ingestion pipeline (images, PDFs, documents), AI-powered condition analysis and recommendations via LLM, semantic search using vector embeddings, and an admin dashboard with invite-only authentication.
• AI Podcast Generator (Python, Vue.js, FastAPI, LangChain): Document-to-podcast conversion tool using GPT-4 for dialog generation and multiple TTS providers (OpenAI, FishAudio) with multi-voice support; features a two-column web editor, voice assignment per speaker, and audio concatenation.
• AI Recruiter Agent (Python, FastAPI, Azure OpenAI, Docker): AI-powered interview management system with automated question generation, dual-tier candidate evaluation (AI + rule-based), semantic candidate matching, multi-channel communication, and asynchronous queue processing for enterprise system integration.
• AI Document Processing System (Python, Next.js, FastAPI, LangChain, PostgreSQL, MinIO, Docker): Multi-agent document extraction platform using GPT Vision and search APIs; supports 15+ file formats, customizable JSON schemas, result versioning with diff and rollback, batch processing, DOCX/PDF template rendering, token-based usage metering, and multi-language UI.
• Autonomous Agent Orchestration System (Python, LangGraph, OpenAI, MCP): LangGraph-based autonomous agent with dynamic tool discovery, MCP server integration, and specialized sub-agents; features configuration-driven tool management (60+ tools), dual LLM support (cloud and local fallback), and a pluggable architecture for file operations, data analysis, web automation, and DevOps workflows.

Publications

• Sequence Alignment by Radial Evaluation of Local Interactions. DOI: 10.2174/1574893613666180130143055. Proposed a novel radial distance metric to generate robust guide trees for MSA; demonstrated statistically superior SP/CS scores vs. MUSCLE/Clustal Omega guide trees on SABRE & PREFAB.
• Modeling the Turning Speed and Car Following Behaviors of Autonomous Vehicles in a Virtual World. Ing. invest. y tecnol. (2015) 16(3):391–405. ISSN 1405-7743. (CARRILLO-GONZALEZ, ARAMBURO-LIZARRAGA, ORTEGA-MAGANA)
• Design of a distributed-control serial bus architecture with dynamical connection among channels ("Zinbus"). Scientific figure & latency analysis: ResearchGate.
• Scientific Computing in Gravitation. Proceedings of the 1st International Supercomputing Conference in México (ISUM 2010), Guadalajara — initial work with CACTUS on UdeG VIRGO cluster.
• Electroencephalographic high-performance cloud data processing system. Proceedings of ISUM 2013, Guadalajara — online cluster processing platform for EEG datasets with significant runtime reductions vs. standalone execution.