Applied AI Engineer — European Central Bank, Frankfurt

AlbertoBarnabò

I take LLM pipelines and agentic applications to production at the European Central Bank — and, on my own time, train small models and publish the evidence.

Selected work Open models

Portrait of Alberto Barnabò — Frankfurt am Main

12,383
model downloads on Hugging Face: 10 / 4
open models / datasets: 51★
stars on lazy-cat: 35
public repositories on GitHub

Live from the Hugging Face and GitHub APIs, updated daily

01 — Selected work

Six things I built, and the numbers behind them.

nDCG@10 on Italian retrieval: bge-m3-italian versus BAAI/bge-m3 on short passages and long documents

01 — Retrieval · Italian

+0.055

nDCG@10 over bge-m3 — long documents, hybrid

bge-m3-italian

Italian retrieval fine-tuned from bge-m3 — a drop-in replacement that beats the base model in hybrid search.

Same 1024 dimensions, same 8192-token context, all three heads — dense, sparse and ColBERT — trained and shipped, so it replaces bge-m3 without changing a line of code. Measured against the base with identical code and a 10,000-sample paired bootstrap on mMARCO-it and MLDR-it; registered on MTEB; ONNX fp32 and int8 builds for Text Embeddings Inference and fastembed, checked against the PyTorch weights.

dense + sparse + ColBERT · mMARCO-it · MLDR-it · MTEB · ONNX fp32 / int8 · right answer at #1: 62.5% → 66%

sentence-transformers · bge-m3 · FlagEmbedding · ONNX

Hugging Face ↗562 downloads

Illustration: a laptop searching a product catalogue with no cloud in the loop

02 — Retrieval

2,100 q/s

retrieval on a laptop CPU

E-commerce product search

Self-hosted semantic search that runs on CPU, with no per-query fees.

A retriever and a cross-encoder reranker fine-tuned on Amazon’s ESCI data, so that a query like “pan that doesnt stick eggs” finds the right product. Both run on CPU with no per-query fees; the repository is the whole factory — every training script, eval and chart. Five open artifacts, including a 31 MB static model for edge deployment.

nDCG@10 0.748 held-out · 45 ms rerank · 5 open artifacts · $2.20 total GPU spend

sentence-transformers · cross-encoder · ONNX · model2vec

GitHub ↗Hugging Face ↗Demo ↗2,325 downloads

Fiduciary’s mascot: an owl in a suit at a laptop

03 — Fine-tuning

9,209

downloads on Hugging Face

Fiduciary

A local-first financial advisor: Qwen3-4B, LoRA, and tool use on Apple Silicon.

A LoRA fine-tune of Qwen3-4B-Instruct on MLX that speaks like a senior personal-finance advisor, wrapped in a small agent loop that reads a portfolio file from disk and calls live price and news tools. Nothing leaves the machine. Published as fused 4-bit weights, a 56 MB adapter, and GGUF for Ollama and LM Studio.

Qwen3-4B · LoRA on MLX · 4-bit · ~45 min training on a 16 GB M-series · 3 published formats

MLX · LoRA · Qwen3 · function calling · GGUF

GitHub ↗Hugging Face ↗

Field-F1 · 85 image documents

Qwen2.5‑VL, inaugural run

Document type	Docs	Field-F1	Best model
scontrino	6	0.881	Qwen2.5‑VL‑7B
modello 730	1	0.800	Qwen2.5‑VL‑7B
fattura (PDF)	9	0.698	Qwen2.5‑VL‑7B
bolletta	20	0.560	Qwen2.5‑VL‑7B
busta paga	38	0.517	Qwen2.5‑VL‑7B
CU	2	0.444	Qwen2.5‑VL‑3B
F24	9	0.343	Qwen2.5‑VL‑7B

Exactly one document in 255 runs was extracted perfectly. Italian document AI is not solved.

04 — Document AI

104

typed fields, from 467 observed

BurocrazIA

The first benchmark for Italian document AI.

Invoices, payslips, F24 tax forms, utility bills, receipts, CU and 730 — a domain with no public dataset, benchmark or eval until now. A 104-field typed schema curated from 467 fields observed on real documents; 204 gold-annotated documents (119 derived by construction from FatturaPA XML, 85 annotated and independently re-verified); a hallucination-counting scorer; and an inaugural leaderboard with Qwen2.5-VL. Italian document AI is not solved.

204 gold documents · 3,805 fields re-verified · 1 perfect run in 255 · best payslip F1 0.517

document AI · VLM evaluation · KIE · Qwen2.5‑VL

GitHub ↗Dataset ↗Demo ↗254 downloads

Five synthetic thermal receipts from the US, UK, Germany, Italy and France

05 — Synthetic data

32,000

receipts, five locales

Synthetic receipts for OCR

32,000 receipts whose ground truth is exact by construction.

Public receipt datasets are small, single-locale and labelled by humans who make mistakes. This generator renders thermal receipts across five locales, then produces a photo-degraded twin of each — homography, uneven lighting, thermal fade, JPEG grunge — with every word box mapped through the same transform. Arithmetic is re-checked per receipt and anything that does not add up is rejected.

US · UK · DE · IT · FR · 4.3 GB on the Hub · pixel-exact word boxes · structured KIE fields

OCR · synthetic data · Pillow · PyArrow

GitHub ↗Dataset ↗1,911 downloads

06 — Agent tooling

18.6×

fewer tokens across 17 tasks

lazy-cat

Claude Code skills that stop the agent from over-working.

Two skills that fire at the only two moments that matter: before choosing an approach (is there an API, a package, a one-liner?) and before writing each block (did anyone ask for this?). Measured across 17 benchmark tasks under three conditions each, the same outcomes cost 4,762 tokens instead of 88,655.

88,655 → 4,762 tokens · 17 benchmark tasks · 3 conditions each · 2 skills

Claude Code · agent tooling

GitHub ↗51 stars

02 — Open on Hugging Face

Every model and dataset, with live downloads.

huggingface.co/albertobarnabo

Name	Kind	Task	Base	Downloads
fiduciary-qwen3-4b	model	financial advisor · text generation	Qwen3-4B, 4-bit MLX	8,753
synthetic-receipts-ocr	dataset	OCR / KIE · 32k receipts	generated, 4.3 GB	1,911
ecommerce-product-search-embeddings	model	product retrieval · embeddings	bge-small, 33M	1,191
prompted_tabfact	dataset	table fact-checking · prompts	TabFact	989
bge-m3-italian	model	Italian retrieval · dense + sparse + ColBERT	bge-m3, 568M	562
ecommerce-product-search-reranker	model	product reranking · cross-encoder	MiniLM, 33M	560
fiduciary-qwen3-4b-GGUF	model	financial advisor · llama.cpp / Ollama	Qwen3-4B, GGUF	456
ecommerce-product-search-embeddings-base	model	product retrieval · embeddings	bge-base, 109M	331
burocrazia	dataset	Italian document-AI benchmark	204 gold docs	254
ecommerce-product-search-embeddings-static	model	product retrieval · static embeddings	model2vec, 31 MB	243
esci-product-search-pairs	dataset	retrieval training pairs	Amazon ESCI	201
fiduciary-qwen3-4b-lora	model	financial advisor · adapter	LoRA, 56 MB	0
scenesmith-qwen3-4b	model	Manim code generation · adapter	Qwen3-4B, LoRA	0
Live · all-time downloads · revalidated daily				15,738

03 — Tools & experiments

The long tail.

OpalZero

Rust

A multi-agent engine in Rust: one plain-English intent in, a typed MissionState out over SSE.

GitHub ↗Site ↗

Italy in 2025 as three columns of dots: workers, pensioners, everyone else

Italian Bankruptcy Forecast

Simulation

Italy from 2025 to 2050, one dot per thousand people — workers, pensioners and everyone else — animated from Eurostat, Istat and INPS data. The watchable face of a larger agent-based simulation.

Demo ↗

tokentropy

Uncertainty

Claim-level uncertainty scores for streaming LLM output, from top-k log-probabilities alone — no second model.

GitHub ↗

professor

Claude skill

Designs an explanation before anything is produced — base → vary → name — with eleven evidence-tagged gates.

GitHub ↗

A Manim keyframe: a 3D surface under a moving camera

manim-craft

Claude skill

Manim craft for LLM-generated animations, plus a blind A/B that showed craft alone doesn’t improve teaching.

GitHub ↗

limn

Python

Zero-dependency charting that accepts the CSV finance actually exported and returns publication-grade SVG.

GitHub ↗

bibliotech

Agent memory

Dated decision records for an agent, so decisions survive across sessions instead of being re-derived.

GitHub ↗

A desk with books, a manuscript and a magnifying glass

ocr-sota

OCR

An automatically ranked index of open-source OCR engines, document parsers and vision-language models.

GitHub ↗

MenuGen

Next.js

A menu spreadsheet in, batch AI food photography out — with the cost shown before you generate.

GitHub ↗

plasmodio

Simulation

A slime mould redraws Italy. Agent-based simulation in the browser.

GitHub ↗Demo ↗

orbis

Swift

A native globe for the countries you’ve lived in, visited, passed through.

GitHub ↗

InterVous

LLM

Parses a job application out of a posting URL.

Live ↗

About

Italian, in Frankfurt, building AI that has to work on Monday.

I studied computer science and engineering at Politecnico di Milano and finished with a double master’s degree at Xi’an Jiaotong University, where I spent two years and wrote my thesis on language models.

Since 2025 I’ve been an applied AI engineer on the European Central Bank’s internal AI team, where I take LLM pipelines and agentic applications from proof of concept to production: document understanding, retrieval and search over large collections, and the unglamorous parts that make a pipeline hold up — evaluation sets before implementation, cost and latency, error handling and retries. I also present and defend those choices to the business teams that use them.

On my own time I train small models on a laptop, publish what comes out, and write down what I learned — including the results that didn’t work. The projects above are that habit in public.

2025 —

Applied AI Engineer

European Central Bank, Frankfurt

Internal AI team. I take LLM pipelines and agentic applications from proof of concept to production — document understanding, retrieval and search over large internal collections — and build the evaluation sets and benchmarks that decide what ships. Day to day: containerised services on cloud infrastructure, cost and latency, error handling and retries, and explaining and defending technical choices to the business teams that use them.

2022 – 2024

AI Researcher

Xi’an Jiaotong University

NLP and large language models for fact verification over tables — the master’s thesis.

2021 – 2024

M.Sc. Computer Science & Engineering

Politecnico di Milano · Xi’an Jiaotong University

Double-degree programme: one year in Milan, two in Xi’an.

2017 – 2021

B.Sc. Computer Science & Engineering

Politecnico di Milano

Research · Master’s thesis · Xi’an Jiaotong University & Politecnico di Milano · 2024

Large language models for fact-checking over tabular data

How well can a language model read a table and decide whether a claim about it is true? The thesis studies LLM behaviour on TabFact and FEVEROUS and how prompting strategy changes accuracy. The prompted TabFact variants are published as a dataset.

Paper ↗Dataset ↗GitHub ↗

Alberto outside the European Central Bank tower in Frankfurt — Frankfurt — outside the ECB

Alberto at Hua Shan, Shaanxi, among red prayer ribbons — Shaanxi — Hua Shan, during the Xi’an years

Shaanxi — Hua Shan, during the Xi’an years

Alberto at his master’s graduation at Xi’an Jiaotong University — Xi’an — master’s graduation, 2024