I build LLM-powered applications designed to work beyond the demo: multi-agent pipelines with validated structured outputs, RAG systems grounded in real documents, full-stack AI products, and end-to-end observability that shows how every answer was produced.

• 🤖 AI Engineer — LLMs, RAG, agents and LLMOps with Python, LangChain, OpenAI and Claude.
• 🧩 Software Engineer — complete products: architecture, integration, deployment and the end user. Former Full Stack instructor at Henry.
• 📐 Mathematician (Universidad de Antioquia) — I approach problems with rigor, from backpropagation to system design.

🔭 Open to AI Engineer, LLM Engineer and Junior ML Engineer roles.

Compares an original contract against its amendment, both provided as scanned images, and returns a validated JSON report of the changed sections, the legal topics affected and a summary of the changes. GPT-4o Vision transcribes each contract, a Contextualization Agent maps how sections correspond, and an Extraction Agent classifies every addition, deletion and modification. Double validation (structured outputs + Pydantic) with retries, and every run traced in Langfuse. Built for a scenario where a compliance team spent 40+ hours a week on this by hand.
GPT-4o Vision LangChain Pydantic Langfuse tenacity Python

Full-stack Claude pipeline that turns a feature-length screenplay (PDF / TXT / Fountain) into scene-by-scene cinematic video prompts, each paired with reference images matched automatically from Google Drive. A world bible keeps characters and locations consistent across ~400 prompts; a tiered matcher (character → location → embeddings) attaches assets. Multi-user auth with scene-level task assignment, prompts streamed live to the browser, and Inngest monitoring to retry failures on the fly. ~$4 in API costs per feature film thanks to prompt caching.
Anthropic Claude Next.js TypeScript Inngest Supabase Prisma Google Drive API

Classifies customer queries by department (HR, IT, Finance) with structured outputs and routes them to a domain-specific RAG agent, with a human fallback for low-confidence cases. An Evaluator Agent (LLM-as-a-judge) scores relevance, completeness and accuracy, all inside a single Langfuse trace per query.
LangChain Chroma Pydantic Langfuse OpenAI

Streamlit research assistant that goes from a topic to a well-cited report: asks clarifying questions, plans targeted searches, iterates with OpenAI's Responses API web_search tool until the goal is met, and exports the report as Markdown, TXT or PDF.
OpenAI Responses API Streamlit ReportLab Python

Serverless recommender built on AWS Bedrock with Claude Sonnet. A Supervisor Agent interprets intent and delegates to specialized sub-agents (hotels vs. Airbnb, fine dining vs. casual), prompting for missing context and filtering S3 data in real time through Lambda.
AWS Bedrock Claude AWS Lambda S3 Python

Upload PDF, DOCX or TXT files and ask questions about them. Content is chunked, embedded and indexed in FAISS; the most relevant chunks are passed to GPT for context-aware answers.
OpenAI LangChain FAISS Streamlit

AI & LLM

Machine Learning

Full-stack

Cloud & Dev

alejandro = {
    "roles": ["AI Engineer", "Software Engineer", "Mathematician"],
    "education": [
        "Mathematician — Universidad de Antioquia",
        "AI Engineering — Henry",
        "Full Stack Web Development — Henry",
    ],
    "experience": ["Full Stack Instructor @ Henry", "E-learning Analyst @ Henry", "Math & Physics Teacher"],
    "how_i_work": ["structured problem-solving", "clear technical docs", "product mindset", "self-driven learning"],
    "hobbies": ["art", "tech", "science", "languages"],
}