About

I’m Juan Francisco Lebrero, an LLM research engineer at Mercado Libre and a builder who ships products end to end, from the model to the invoice.

By day I work on Mercado Libre’s LLM Science team: in-house LLM training and post-training, across continued pretraining, supervised fine-tuning, data curation, and evaluation benchmarks. On my own time I run an AI consultancy and ship products: a commerce SaaS that bills Argentine businesses, an autonomous prospection agent, and open-source tools that people actually use. Most people in AI sit on one side of a line. Researchers who can’t ship, or product engineers who treat the model as a black box. I work on both.

What I’m working on

Mercado Libre, LLM Science. In-house LLM training and post-training: continued pretraining (CPT), supervised fine-tuning (SFT), and domain adaptation for large-scale commerce, plus evaluation benchmarks for instruction-following, reasoning, regression quality and reliability across model iterations.

AWAM (awam.lat), my own venture: the execution operating system for Argentine fashion brands. An AI-first commerce platform that runs the whole loop, from a customer’s WhatsApp message to recommendation, checkout, payment, electronic invoice, stock, and lost-demand recovery. Six specialized AI runtimes over a durable, multi-tenant core.

sesgo.ai (sesgo.ai), my AI consultancy: machine learning, data science, AI agents, data engineering and MLOps for teams across LatAm and the US, built for measurable ROI.

Research at LiNAR (UdeSA). A vision foundation model for non-invasive embryo assessment in IVF: JEPA-style self-supervised representation learning over time-lapse imaging, predicting viability without a biopsy.

Selected past work

SOFLEX (Lead Data Scientist, 2025 to 2026). Led an AI team building a low-latency intelligence platform over 200M+ emergency reports, with real-time situational awareness in under 3 seconds, plus domain-tuned LLM and ASR pipelines for Argentine Spanish.

Papelera San Andrés de Giles. A multimodal SKU-verification pipeline (VLM + OCR) at 100% accuracy on deployment tests, running 600 pallets/day.

UK University (México). AI lead; cut MVP cycles from a week to a day, +17% enrollment, US$108k/year saved.

How I work

I’m cost-obsessed and local-first. I benchmark before I have opinions, I write down my reasoning as decision records, and I dogfood my own tools. I measure impact in numbers ($0.06 per company prospected, US$108k/year saved, 100% SKU accuracy on 600 pallets/day) because impact you can’t measure is impact you can’t defend. I build a lot of agentic infrastructure, like memory, orchestration, concurrency and recovery, because that’s the scarce skill as teams go from one agent to a fleet.

Background

AI Engineering at Universidad de San Andrés (95% merit scholarship, 8.67/10 GPA, top 10 in the cohort), two-time ICPC Latin America regional qualifier, hackathon and game-jam winner. I move fluently between Python, C/C++/CUDA and TypeScript.

→ See what I’ve built on the Projects page.

Contact