A retrieval agent trained on everything I've built. It answers, and it cites the source.
Ask Santi
live
>
reading the sources...
scroll ↓
01SELECTED WORK
Built to be taken apart.
Four systems, each documented like a component: what it is, what it does, and the facts that hold up.
PART 01 / MASTERY-ENGINE
2024 to now
Personal build
Mastery Engine
A multi-agent LLM system where a controller orchestrates handoffs between specialist agents, with tool execution, structured outputs, context management and a spaced-repetition scheduler. Built to run and adapt, not to demo.
Multi-agent
orchestrated handoffs
Spaced-rep
scheduler + tool calls
TypeScriptOpenAISQLiteDocker
PART 02 / BATTWIN-THESIS
2026
Master thesis / KTH
Federated Fault Prediction
My master thesis as a Research Intern at KTH Stockholm, inside the EU Horizon Europe BATTwin project. Machine learning for fault and defect prediction in battery-cell manufacturing, using federated learning so a shared model trains without any plant centralizing its data.
A federated model trained across decentralized medical-imaging data, without ever bringing that data into one place. Built on Flower and PyTorch across multiple clients, reaching roughly 96 to 97% global accuracy.
96-97%
global accuracy
Flower
PyTorch · multi-client
Federated learningFlowerMedical imagingPrivacy
PART 04 / SMART-RACKET
2024
On-device
Smart Padel Racket
Embedded AI running fully on a microcontroller: real-time inference in C/C++ off an IMU with TensorFlow Lite Micro, paired with an iOS dashboard. All classification happens on the racket, with no cloud round trip.
Real-time
on-device inference
TFLite Micro
IMU · iOS dashboard
Embedded C/C++TensorFlow Lite MicroIMUiOS
02WRITING
Notes from the workbench.
Technical posts on agents, retrieval and shipping ML to constrained hardware.
Bachelor thesis at 9.5/10: ML fault detection in EV power converters, embedded for real-time inference. Where the habit started: take the system apart until the math is obvious.
Fides Electrónica / Sant Cugat
Embedded & IoT trainee
Firmware in C/C++ on ESP32, STM and Renesas under tight memory and power budgets. Repaired 15+ PCBs. Learned hardware by its failures.
Tecnun, U. Navarra / San Sebastián
MSc Data Analysis in Engineering
MADI, GPA 9.0. Machine learning, deep learning, NLP, reinforcement learning and AI for Industry 4.0.
KTH / Stockholm
Master thesis, EU Horizon BATTwin
Research Intern applying federated learning to fault prediction in battery-cell manufacturing.
Now
Agentic AI & ML engineer
Multi-agent systems, retrieval and on-device ML.
OFF THE CLOCK
I understand things by taking them apart. That is the same move whether it is a multi-agent system or a carburettor from a classic motorcycle in my garage: strip it to parts, name every part, put it back so it runs better than before.
Outside the terminal that instinct goes into restoring old bikes, padel, the gym and long runs. I also refereed basketball for four seasons, which is really just making fast calls under pressure with everyone watching. The mindset does not switch off; it just changes tools.