I build LLM-agent skills and RAG pipelines that run in production and save engineering teams real hours.
I'm an AI Developer at Mandü HR with 3 years of experience and a QA → Full Stack → AI progression inside the same production HR tech.
I'm part of the team that maintains the engineering area's internal LLM-agent toolkit, and I own the QA vertical: I design, maintain and evolve every QA skill the team consumes. The RAG pipeline I designed cut test-case creation time by 70.9% (50h → 14.55h per sprint), and I refactored it into the modular skills I maintain today.
My focus is maintainability: impact that holds up sprint after sprint, not just in a demo.
- LLM agents & RAG in production — Jira/Confluence ingestion, pgvector + ChromaDB, OpenAI & Anthropic APIs, MCP integrations, context engineering
- QA automation with LLMs — test-case generation from sprint docs, Jira upload, endpoint → Postman mapping, Playwright E2E generation with automatic evidence
- Multi-agent orchestration — LangGraph state machines with review-retry loops (see the public repos below)
- Full-stack foundation — NestJS, React, Angular, Laravel, .NET; SQL/NoSQL; Azure, Docker
| Problem | Migrating a legacy codebase (Java, PHP, TypeScript, Python, COBOL) to a clean modern architecture is months of manual, error-prone rewriting. |
| Solution | A 6-stage LangGraph state machine — Ingest → Parse → Index → Plan → Generate → Review — that reverse-engineers a repo, embeds it into PostgreSQL + pgvector for RAG, maps bounded contexts to NestJS modules and Angular features, and generates the project with a 4-pass review loop that retries on critical findings. |
| Stack | TypeScript LangGraph RAG pgvector OpenAI API NestJS Angular Clean Architecture |
| Result | Real migration included in the repo: a ~40-file IoT sensor platform → 37 files across 3 NestJS modules in 4 min 39 s, with a migration quality report. |
| Repo | DiegoDeLaFlor/Legacy-to-Modern-Architect |
| Problem | Fixing a bug ticket well means three different jobs: find the root cause, propose the minimal fix, and try to break it. One agent doing all three tends to skip the last one. |
| Solution | An MVP of three cooperating agents on LangGraph — Investigator (RCA, locates suspect files), Programmer (minimal fix with impact assessment), Verifier (runs tests and tries to break the change) — driven from a local CLI with a JSON ticket. |
| Stack | Python LangGraph CLI pytest |
| Result | Console report with RCA, fix proposal, PASS/FAIL evidence and logs for every ticket. |
| Repo | DiegoDeLaFlor/Agent-for-bug-fixing |
| Problem | Teams have no cheap way to turn a GitHub repo into actionable architecture and quality insights. |
| Solution | SaaS MVP: ASP.NET Core backend (DDD + Clean Architecture) with GitHub OAuth and repo cloning, a Roslyn AST analyzer, a FastAPI AI engine that generates insights, and a React dashboard. Orchestrated with Docker Compose. |
| Stack | C# / .NET 9 Roslyn DDD Clean Architecture FastAPI Python React Docker |
| Status | Working MVP (analyze → view issues & insights). Next: EF Core persistence, multi-language AST. |
| Repo | DiegoDeLaFlor/DevInsight-Platform |
| Problem | Writing test cases by hand from Jira user stories and Confluence specs took ~50 hours per sprint. |
| Solution | RAG pipeline over Jira + Confluence (pgvector + ChromaDB, OpenAI/Anthropic API), refactored into modular skills inside the internal LLM-agent toolkit: test-case generation, automatic Jira upload, endpoint → Postman mapping (happy & unhappy paths), and Playwright E2E generation that runs each case and captures evidence. |
| Impact | 70.9% less time creating test cases (50h → 14.55h per sprint), sustained in production and used daily by QA teams. |
| Problem | Coffee farmers lack data-driven guidance based on real field conditions. |
| Solution | End-to-end research project: ESP32 sensors (humidity, NPK, rain) → Spring Boot edge backend → .NET on Azure → REST API serving predictions. |
| Result | 3 models evaluated (Random Forest, XGBoost, CatBoost) on ~9,983 samples; Random Forest selected for production. |
| Solution | n8n flow chaining OpenAI (script) → Stability AI (images) → ElevenLabs (voice) → Google Sheets/Drive (queue & assets). |
| Result | 20+ shorts per month generated with no manual intervention. |
|
|
|
Note: Public metrics reflect open-source and personal projects. Most of my production work lives in private organizational repositories under confidentiality agreements.
Earlier work from my degree, kept public for reference: Four Dreams (DDD microservices on Azure, .NET) · AdventureHub (Flutter + NestJS) · Diseño de Experimentos (Jenkins CI/CD, Cucumber BDD, JMeter) · Digital UX — DermApp (React/Angular healthtech)
I also run a YouTube channel on mystery and narrative-driven discovery. Structuring a story from research to final cut is the same discipline as designing a system: know the audience, cut what doesn't serve the outcome.
I'm interested in:
- LLM agents and RAG systems that ship to production and stay maintainable
- Developer tooling that removes real hours from engineering workflows
- Open-source work on agent orchestration and QA automation
Built with passion for engineering excellence. Dark mode optimized. © 2026