I’m Sebastian, an ML engineer from Bogotá and a biomedical engineer from Universidad de los Andes. At Provectus I ship LLM, RAG, and agentic systems on AWS that hold up in production.
I build evaluation harnesses for LLM extraction, agents that review actuarial reports, LLM judges calibrated with DSPy, retrieval over S3 Vectors and OpenSearch, sign-language models from EMG and video, clinical ML that helps doctors agree.
Machine Learning Engineer with hands-on experience shipping production LLM, RAG, and agentic systems on AWS for document-intensive enterprise pipelines. MSc in Biomedical Engineering with research on multimodal data collection, computer vision, and deep learning.
Comfortable working end-to-end across ML backend, serverless infrastructure (AWS Lambda, Step Functions, SAM), and full-stack delivery. Focused on turning frontier AI methods — LLM evaluation, retrieval, and human-in-the-loop workflows — into reliable, observable, cost-aware production systems.
Outside client work I build with Claude every day: a personal Android app in Kotlin and Jetpack Compose with a FastAPI assistant on Lambda that calls the Anthropic API behind spend caps, output checks and evals, and Next.js sites with Resend contact flows. Claude Code is how I work, with subagents, git worktrees, skills and MCP, and I’ve run internal workshops on it.
Building LLM document-intelligence systems for enterprise insurance clients, end to end: extraction pipelines, agents, evaluation, and the web apps around them, on AWS.
Part-time agentic AI and full-stack development. Designed and shipped agentic workflows that turn unstructured PDFs into validated, structured records inside the product.
Led laboratory sessions on machine learning fundamentals: optimization techniques, linear and logistic regression, analytical OLS solutions, hyperparameter tuning, and neural network architectures.
“From Expert Judgment to Autonomous Optimization: Encoding Human Expertise into LLM Judges with DSPy”
How we evaluated an LLM pipeline for reinsurance contracts: five domain experts’ reviews distilled with DSPy into a panel of calibrated LLM judges, within 10% of expert consensus on every dimension, then used as the fitness function for autonomous prompt optimization with MIPROv2 and GEPA.
Thesis: Machine Learning Analysis of Colombian Sign Language: Recognition, Classification, and Biomechanical Characterization.
Coursework: Machine Learning for Engineering · Reinforcement Learning · Analysis & Processing of Medical Images
GPA 4.74 / 5.00Coursework: Data Structures & Algorithms · Scientific Programming · Signal Processing · Neuroscience · Neuroanatomy
GPA 4.08 / 5.00