I build and break intelligent systems.
ML engineer with 8+ years shipping production AI, currently a Senior AI/ML Engineer at Nequi (Grupo Bancolombia), where I pre-train foundation models on customer transaction sequences. Off the clock, the center of gravity is research: agent security and red teaming, LLM evaluation, and mathematical reasoning, mostly played out in the open through competitive ML. Kaggle Competitions Expert, ranked 1,094 of 209,674.
WHOAMI
I spent eight years learning how ML systems fail in production (drift, silent data bugs, ungrounded generations) and building the pipelines, evals, and observability that catch it. That engineering discipline is now my research method: every open question I work on gets a documented brief, a scoring model analysis, an experiment ledger, and a reproducible pipeline.
Today that method points at the questions I find hardest and most consequential: how agentic systems break (red teaming, multi-step tool attacks), how to measure them honestly (LLM-as-judge, offline evals wired into CI), and how they reason (single-primitive normal forms for mathematics, program induction). Most of it plays out on Kaggle, where the leaderboard keeps me honest.
Also: TEDx speaker on AI, international hackathon winner (Baubap ML Hackathon, 1st place), trilingual (ES · EN · PT), and mentor to 13+ engineers.
class SebastianMateus:
role = "ML Engineer → AI Researcher"
base = "Bucaramanga, CO · remote (UTC-5)"
research = ["agent security", "LLM evaluation",
"mathematical reasoning"]
stack = ["Python", "PyTorch", "LangGraph", "SQL",
"dbt", "BigQuery", "Docker", "Terraform"]
oss = ["eml-translator", "barrio-mapper", "realh"]
languages = ["es", "en", "pt"]
kaggle = "Expert · rank 1094 / 209674"
ships = TrueRESEARCH
Independent research run like production engineering: every line of work below has a public artifact, a reproducible pipeline, and an experiment ledger behind it.
Algorithmic red teaming: writing attack algorithms that surface reproducible multi-step failures in tool-using agents (GPT-OSS, Gemma) under public and private guardrails, inside a deterministic, replay-validated sandbox. Built a free-tier evaluation harness that estimates true per-model scores without spending scarce submissions, separating learning from scoring.
Competition ↗ARC is the benchmark built to resist memorization: novel reasoning from a handful of examples. Finished 175th of 1,455 in ARC Prize 2025 while leading the competitive ML team I founded at Factored, and I'm currently entered in both ARC-AGI-2 and ARC-AGI-3 for the 2026 prize.
ARC Prize 2025 ↗ ARC Prize 2026 ↗ Write-up ↗
Independent Python implementation of Odrzywołek's exp-minus-log primitive, the NAND gate of
continuous mathematics, released as the eml-translator library. On top of it, an agent
research POC: a 21-problem benchmark (100% solved), seven EML tools exposed via Claude's tool-use
API, and an 8-subagent A/B investigation of four Riemann-Hypothesis-equivalent criteria. The POC
also reports where the tools don't pay: on calculus and algebra the agents route around
EML to sympy.
Copy-move forgery detection and segmentation in scientific figures, a direct attack on research fraud. Bronze medal, 128th of 1,564 teams in the Recod.ai/LUC challenge, working from a public baseline I rebuilt around a different segmentation head. Related scientific ML work includes exoplanet signal recovery (NeurIPS Ariel Data Challenge, 184/860) and ECG digitization from paper printouts (PhysioNet, 250/1,424).
Competition ↗ Write-up ↗Detection and tracking of zebrafish embryo cells through 3D space and time. Reverse-engineered the official metric into an offline scorer that runs without images, enabling fast local iteration; current focus is cell division detection, the largest unclaimed scoring lever.
Competition ↗Detecting papyrus surfaces inside CT scans of carbonized Herculaneum scrolls. I built a 3D hysteresis fusion: high-confidence voxels seed morphological propagation inside a permissive two-model union mask. Of everything I've submitted, this is the one whose private score beat its public score, which is the only real evidence a method generalizes rather than fits the leaderboard.
Competition ↗ Write-up ↗Every project above is run as a research program, not a leaderboard sprint: a written competition brief, a first-principles analysis of the scoring model, baselines with provenance, an experiment ledger, and submission pipelines that never spend a run to learn what a free evaluation can tell me.
CAREER
Eight years from analytics to production ML to leading teams: the engineering backbone behind the research.
Senior AI/ML Engineer
Competitive ML Team Lead & Founder
Machine Learning Engineer (L2 → L3)
Senior Data Scientist
COMPETITIVE ML
Competition is my favorite forcing function: a frozen metric, a hard deadline, and no partial credit for good intentions. Twenty-seven entered, seven running right now.
Read this before the numbers. Both medals came from public notebooks I ran, not methods I invented — the silver was a verbatim copy, similarity 1.000. All 19 finished competitions are classified in a public provenance ledger: 5 substantially original, 5 derived, 6 verbatim copies, 3 teammate-authored, every upstream author named. It was built by diffing my final submissions against the notebooks they came from.
Running a strong public notebook is legal on Kaggle and teaches you very little. These are the entries where the approach was my own, which is the only part of a leaderboard worth showing.
Each links to a full technical write-up. Every competition I finish gets one, with the method as actually submitted and credit to any public work it built on.
Ranked 1,094 of 209,674 Kaggle competitors, with a silver and two bronze medals, and nine top 20% finishes. Twenty-seven competitions entered across tabular, vision, NLP, audio, optimization, and simulation tracks.
Agent security, embryo cell tracking, Pokémon TCG battle agents, knee abnormality detection, the Kaggriculture challenge, and both ARC-AGI-2 and ARC-AGI-3, each run with a written brief, an offline scorer, and an experiment ledger.
Won the neural networks category of the Mexican fintech's international hackathon in CDMX: fraud detection over 1K+ tabular features at 90%+ accuracy.
I write about data science, AI, and the scientific method on Medium, and I gave a TEDx talk on how AI reshapes the way we solve problems.
TOOLKIT
CONTACT
If you're building (or breaking) intelligent systems and want to compare notes, the fastest way to reach me is LinkedIn.