Hi, I'm German 👋

AI Application Engineer · LLM agents for business systems.

I build software where business systems meet LLMs. My foundation is ERP: years of Odoo development and migrations, from custom modules and third-party integrations to the infrastructure they run on. That's the part most AI projects underestimate: the data model, the workflows, and what breaks in production. Now I build with LLM agents every day, spec-first and with the model kept on a short leash: it interprets messy input, code does the math and enforces the rules.

🧭 What I do

• Implement and migrate Odoo end to end: custom modules, integrations, fiscal localizations, data migrations.
• Run the infrastructure around it on AWS, with CI on GitHub Actions and pre-commit standards.
• Automate the systems that stay: browser automation where APIs fail, bots that turn chat workflows into structured ERP data.
• Build with LLM agents daily: subagents, persistent project memory, one setup versioned across machines.

📈 A few things I have shipped

• 20+ end-to-end Odoo migrations (v15–v17) across every core module, plus integrations (Mercado Libre APIs) and fiscal localizations for Mexico, Peru and the US.
• Migrated a 100GB production database by working around a framework-level batch-update limitation that blocked the standard path.
• An AWS browser-automation service (Playwright) that restored bulk CFDI downloads after SAT API changes broke the standard integration.
• A Telegram bot that replaced a WhatsApp payment workflow with structured approvals in Odoo that trigger invoicing automatically.
• Found and fixed a calculation bug in a proprietary commissions module during its migration; the fix went upstream.
• A bug fix in Mozilla's Firefox DevTools MCP (#228): fill_by_uid concatenated old and new text on fields that restore their value on focus, such as React controlled inputs. Reproduced it, made the fill edit the field the way a user does, and added tests that fail without the fix.

🔧 Featured work

• Auditing users and permissions in a legacy ERP: reverse-engineered an undocumented Firebird schema and built a report → review → guarded cleanup cycle for Aspel SAE/BANCO, with a single atomic write path, rehearsal and verified backups.

• AltScore Makers Challenge: 8 of 9 hiring puzzles against a live API in about a day, enough for the CTO interview; I planned each solution and Claude Code wrote the code from that plan.

• Engineering standards for a holding's Odoo rollout: Odoo.sh as the runbot, fast-forward releases, GitHub rulesets designed around the platform's real limits, and a lint gate proven with negative tests, all in place before the first custom module.

• Planning an Odoo agent with an agent council: I built a Claude Code plugin that plans, applies and tests Odoo 20 configuration on demo databases, from the plan of an experimental multi-agent pipeline. The page reports what that plan cost and broke: 30 tasks shipped, 23 defects introduced by the plan and 11 failures prevented, in one build with no second condition to compare, so the question stays open.

• firefox-mcp-launcher: runs Mozilla's Firefox DevTools MCP against a dedicated Firefox instance, alongside the everyday browser. The instance runs Marionette and BiDi and uses a clean profile by default, so the agent never inherits logged-in sessions. Also documents the failure modes I hit, such as a second MCP client locked out of Marionette behind an opaque error, and the launch command that silently starts nothing (fixed upstream in my merged docs PR mozilla/firefox-devtools-mcp#220).

🛠 Daily tools

ERP Odoo (v15–v20), QWeb, OWL
Languages Python, JavaScript, SQL, Bash
Data PostgreSQL
Cloud & delivery AWS (EC2, IAM), Docker, GitHub Actions, pre-commit
Automation Playwright, Telegram Bot API
AI Claude Code, LLM agents

📫 Let's talk

• LinkedIn: German Loredo Becerra