Jordi Granja
Jordi Granja

Jordi Granja

AI/ML Engineer

I just finished a Bachelor's in Artificial Intelligence at UPC (top 5% of cohort), and I've spent the last year and a half building ML systems in production.

Most recently at Nestlé, I built an LLM-powered alert triage pipeline for their global Cyber Security Operations Center: BM25 retrieval, RAG, real-time threat intelligence. It cut alert volume by over 50% with near-zero false negatives, saving 1.4 analyst FTEs per year. Before that, at Strategic Platform, I shipped everything from hybrid retrieval systems and topic modeling pipelines to an A* pathfinding service for orienteering races.

On the research side, I love learning from replicating papers. I've implemented HER, TD3, DDPG, offline RL, and RLHF from scratch, and optimized Vision Transformer inference on the MareNostrum supercomputer with CUDA. I like understanding things at the level where the math and the engineering meet.

I'm looking for my next role. Ideally somewhere applied AI is treated as a first-class engineering problem.

Projects

Hindsight Experience Replay (HER)

Reinforcement Learning • PyTorch • Goal-Conditioned RL

Implementation of Hindsight Experience Replay, the technique that lets sparse-reward RL agents learn from failed episodes by retroactively relabeling goals. Built for tasks where the reward signal is nearly absent.

HER shows an interesting perspective; we can avoid engineering complex, dense reward functions by retroactively learning from failures.

Code →

Actor-Critic Algorithms: DDPG & TD3

Reinforcement Learning • DDPG • TD3 • Continuous Control

Clean replication of two deep RL algorithms for continuous action spaces — Deep Deterministic Policy Gradient and Twin Delayed DDPG. Implemented from scratch following the original papers, tested on standard MuJoCo benchmarks.

TD3's tricks (target policy smoothing, delayed updates, twin critics) are small changes with notable impact.

Code →

Reinforcement Learning from Human Feedback (RLHF)

RLHF • Reward Modeling • PPO • Alignment

Implementation of the RLHF pipeline: supervised fine-tuning, reward model training from preference data, and policy optimization with PPO. Built to understand the mechanics behind how modern LLMs are aligned with human intent.

Code →

Large-Scale ML Pipeline (Big Data Analytics)

Apache Airflow • PostgreSQL • HDFS • Python

End-to-end data engineering and ML pipeline built for large-scale batch processing. Orchestrated with Apache Airflow, storing and querying across PostgreSQL and HDFS. Covers ingestion, formatting, quality testing, and exploitation.

This project made the infrastructure side of ML tangible.

Code →

More projects on GitHub.

Skills & Interests

Technical Skills

Areas of Interest

At the moment, I am most interested in large language models — how they are trained, how they are evaluated, and how to run them efficiently at scale. I find the evaluation side especially interesting: it is hard to know if a model is actually improving at something useful, and I think this is one of the most important open problems in the field.

I also care a lot about the infrastructure side — retrieval systems, agentic pipelines, and making models reliable enough to use in real products. Working at Nestlé and Strategic Platform showed me how different production is from research, and I want to keep working at that intersection.

Currently Learning

I am currently studying scaling laws — what they tell us about how models improve with more data and compute, and where they do not hold. On top of that, I have been reading about agentic AI: how to build systems where a model can use tools, remember things, and complete multi-step tasks reliably.

Writing

I'm planning to write about my projects and the things I learn along the way. Topics will include practical ML implementation, deployment challenges, and interesting problems I encounter. Check back soon!

Contact

The best way to reach me is via email. I'm also active on GitHub and LinkedIn.

Download CV (PDF)