Skip to content
Joaquín León Alderete

Joaquín León Alderete

AI Engineer, Buenos Aires, Argentina

Deep learning research and AI systems that run in production.

I work at the intersection of two things that are usually kept apart: research and systems with users who depend on them. On both sides the question is the same: how to make an AI system trustworthy when something depends on it.

Building
Verbiux, the AI newsroom platform
Now
Lead engineer on a Swedish govtech deployment
Award
Winner of the global Gemma 4 hackathon, USD 10K prize
Research
NLP at ELIAS Lab, satellite segmentation at LINAR

Work

Now

Lead Engineer · Swedish govtech deployment

Won a public tender and lead the development of a system that helps municipal case officers reconstruct unpermitted-construction case files by cross-referencing geodata with information scattered across many sources. Built around traceability and declared uncertainty.

GeodataPublic sector

Nov 2025 - Present

AI Solutions Architect · El Intransigente

Designed and shipped PressMate, a multi-tenant B2B SaaS for digital newsrooms. LLM workflows run as state machines on a stateless FastAPI and LangGraph backend; Redis acts as message broker and pub/sub bus, streaming progress to the UI over SSE. Tenant isolation and role-based access are enforced at the database level; OAuth integrations with WordPress, Meta and Google Analytics 4.

Used by seven outlets. 600+ articles and 1M+ pageviews per day.

PythonFastAPILangGraphRedisNext.jsPostgreSQLDrizzleCloudflare R2

Now

Researcher · ELIAS Lab, Universidad de San Andrés

NLP research on natural language autoencoders.

NLPAutoencoders

Sep 2025 - Present

Research Assistant · LINAR, Universidad de San Andrés

Selected for the Research Assistant Program. Co-lead a project on semantic segmentation of multispectral satellite imagery, end to end: dataset preparation, literature review, experiments and analysis.

Deep learningSegmentation

Now

AI integrations · Aerowise

Intelligent integrations for private pilots.

LLM

Research

DSS-Net

Asymmetric dual-stream segmentation for multispectral satellite imagery

AcceptedJornadas Argentinas de Robótica 2026

Under limited labeled data, a moderate-capacity model matches or beats a geospatial foundation model with an order of magnitude fewer parameters.

Visible and near-infrared bands are processed in separate branches and fused with attention. Architectural adaptation outperformed scaling, at least there. What matters to me is understanding what the model actually uses to decide, instead of assuming it.

Natural language autoencoders

NLP research at ELIAS Lab, Universidad de San Andrés

Ongoing

Current work at ELIAS Lab on autoencoders for natural language: how to compress text into representations that can be decoded back without losing what matters.

Projects

2026

Autonomous robotics in ROS 2

Monte Carlo Localization, an Extended Kalman Filter and six path planners (APF, Dijkstra, A*, Theta*, RRT, RRT*) for a TurtleBot3 over occupancy grids.

PythonROS 2NumPy

2026

Petrocast

Hydrocarbon production forecasting platform.

Python

Education

  • B.Eng. in Artificial Intelligence

    Universidad de San Andrés, 2023 - 2027

  • JavaScript Mastery, Web Development

    Coderhouse, 2022

Awards

  • Winner, global Gemma 4 hackathon

    USD 10,000 prize

  • Paper accepted at Jornadas Argentinas de Robótica 2026

    DSS-Net, multispectral satellite segmentation

Skills

LLM systems
LangGraph, FastAPI, Redis pub/sub, SSE, multi-tenant SaaS, OAuth integrations
Deep learning
Semantic segmentation, multispectral imagery, attention-based fusion
Robotics
ROS 2, EKF, particle filters, motion planning
Web and data
Next.js, TypeScript, PostgreSQL, Drizzle, Cloudflare R2
Core
Python, C, NumPy, OpenCV, Linux, Git

Contact

Based in Buenos Aires, working with teams in Sweden and Argentina. Write to joaquinleondev@gmail.com.