AI & HPC Lead

Agentic news-intelligence engine LLM · Pydantic · event memory

A system that watches trusted financial news sources and tracks stories instead of individual articles. It decides whether each new article continues an existing event or starts a new one, and it's designed to alert only when an event materially changes, not on every headline.

1. Ingest5 sources, every 30 min
2. Prefilterrules, no LLM: drops noise cheaply
3. ExtractLLM → Pydantic-validated JSON: entities, event types, triggers, figures
4. Gatekeep reporting, drop opinion and recaps
5. Event memoryattach or create · 48 h window · score ≥ 0.65
6. Reason & alertagent over working, episodic and reference memory · in progress

• ~24%of articles dropped before any LLM call
• 0.75agreement bar a local model must meet to replace the API model
• 0.53–0.69what DeBERTa, BART and Qwen reached, so the API model stayed

Built as a tick-driven loop (Python, SQLite locally) with replay for testing, designed to move to AWS (ECS Fargate, RDS, pgvector).

Coverage-bounded H3 maps for service planning independent research · open data

Service planning needs to know, for every place, whether people need urban, suburban or rural service. Fine hexagons get that right but produce thousands of zones per city. I developed a rule that coarsens an H3 grid only where coarsening loses (almost) no one. The result is a much smaller map that still serves nearly everyone at their level of need.

• 6European cities
• 19–56%fewer points at 99.9%, held-out cities
• 4open datasets: HRSL 30 m, GHS-SMOD, OSM, Eurostat

Methods were fixed before testing on Milan, Madrid and Paris. At 80–90% coverage the results are mixed: another method is sometimes cheaper. Built spec-first: tests for every requirement, data provenance hashes, and a browser engine parity-tested against the Python code.

Case study: day-ahead hail prediction EU H2020 CYBELE

1. Input6 TB of downscaled weather simulations: ~50k GRIB files, 68k grid points × 1,000+ features, 5 years
2. Parallel I/Omultiprocessing + PyArrow, converted to Parquet on one Marconi100 node
3. Traininggradient-boosted ensemble (XGBoost, LightGBM, AdaBoost), scheduled with SLURM
4. DeliveryDocker container for the project demonstrator

Multi-GPU debugging with RAPIDS and Dask NVIDIA DGX A100

With colleagues, I traced a bug in a Dask-based all-reduce that made jobs fail when scaling from one GPU to several on an NVIDIA DGX. Fixing it made multi-GPU data processing reliable again.

• RAPIDS cuDF
• RAPIDS cuML
• Dask
• 8 × A100 per system

PCA shortcut for virtual drug screening EuroHPC LIGATE · informal contribution

LIGATE is a EuroHPC drug-discovery project, with CINECA in the consortium, that screens very large libraries of molecules by docking them against protein targets. As a side contribution I showed that dimensionality reduction (PCA) recovers the same subset of candidate molecules that the established linear docking-score equation selects. That makes it a different, faster route to the same shortlist. Individual files ran from tens of gigabytes to terabytes, far too large for memory, so I read them from disk in parallel and benchmarked the approach across systems, including a DGX.

1. Docking datatens of GB to TB per file, on disk
2. Parallel readspandas + Linux tooling, chunked and out-of-core
3. PCAreduce the descriptors, rank the candidates
4. Same shortlistmatches the linear scoring equation's candidates
5. Benchmarksacross systems, including DGX

Finding magnetic reconnection in NASA MMS data EU H2020 AIDA

Magnetic reconnection, where field lines in Earth's magnetosphere snap and reconnect, is rare and hard to spot in spacecraft data. I trained models on plasma simulations, where reconnection is labelled, and tested them on real data from NASA's four-spacecraft Magnetospheric Multiscale (MMS) mission. To make that transfer possible, I put both data sets on a common footing: resampled to the same resolution (0.05 d_i) and each trajectory standardised against itself.

Federated statistics for COVID-19 cohorts EU H2020 ORCHESTRA

ORCHESTRA analysed SARS-CoV-2 patient cohorts across Europe, where GDPR often prevents records from leaving a country. I co-developed the federated infrastructure and built a small statistical package for federated analysis. It computes each analysis locally at every national hub and combines only the aggregates centrally, so individual records never move.

Teaching and reviewing Bologna Business School · CINECA · Fortissimo Plus

I taught the Master's in Data Science course at Bologna Business School and ran training on ML and distributed computing for HPC users. In 2024–25 I evaluated European HPC/AI innovation proposals for Fortissimo Plus.

"AI with HPC" series, Towards Data Science, 2022: 1. Parallel data processing · 2. Multi-GPU training with Horovod

Out-of-home network strategy, city by city strategic network planning

Where should a parcel carrier put its pick-up shops and lockers in each city, and how many does it really need? I built the answer from zero, on my own: the data pipelines, the demand model, the optimiser that sets each city's network against coverage targets, and the planning app on top. Once it was in production, scaling it across Europe needed a team. I then led six people to deploy it in seven countries.

• product owner
• architect
• lead developer
• team lead

1. Demandpopulation, socio-economic and map data per city
2. Targetscoverage goals set by area type
3. Candidatesevery feasible shop and locker site
4. Optimisefewest points that meet the targets, at lowest cost
5. Delivercity plan in a planning app and API for operations

• −20%network operating cost
• €1–5Msaved per year, per country
• 6people led for the roll-out
• 7countries in production

Stack: Python, H3, OpenStreetMap and socio-economic data, mathematical optimisation, a Dash planning app and an Azure Functions API.

Demand forecasting and EU AI Hubs forecasting · EU AI Hubs

Forecasting models in production, with the pipeline re-engineered so a computation that took months now runs in minutes. That turned forecasting from a one-off study into something planners can re-run. As technical lead for GLS's EU AI Hubs, I defined best practices and prototyped solutions for adoption at country level.

• <5%MAPE
• €0.5–1Mvalue per year
• EU AI Hubstechnical lead, best practices

One model, many clients system dynamics · DAG architecture

ScenZero lets manufacturing SMEs plan decarbonisation as a dynamic system instead of a static spreadsheet. As co-founder and CTPO, I owned product and technology.

• The model: system dynamics. Stocks (emissions, installed capacity, capital) change through flows (investment, abatement, energy use). Feedback loops and delays mean a measure's effect on cost and profit plays out over years.
• The key decision: one mother simulation instead of a bespoke model per client. A DAG reads each client's profile, resolves which components and dependencies apply, and deploys only that simulation.
• Why it matters: without it, every client needs its own model. That is a consulting job, not a product.
• Hands-on: I turned the Vensim model into a Python/JavaScript simulation engine and built the app.

• Co-founderCTPO: product + technology
• 1 modelmother simulation, assembled per client by a DAG
• 4app modules: business case, inaction, roadmap, explorer

scenzero.com → · live demo →

Running an R&D product group team · direction · partners

I lead a team of data scientists and software engineers across four products (see the table above). My job is to turn research into something a partner can use and pay for.

• Direction: set research directions and decide which prototypes become products
• Architecture: define the system and its validation criteria before the build
• Partners: roadmaps agreed with external partners, such as WIGeoGIS for LocationAI
• Hands-on: I still design algorithms, for example LocationAI's adaptive H3 grid and its agentic footfall simulation

LocationAI press release (German) → · grydnova.com →

GrydNova: planning local grids for the energy transition electricity grids · in market

Rooftop solar, EV charging and heat pumps are putting pressure on low- and medium-voltage grids. GrydNova helps German grid operators decide where and when to reinforce. I provide the product and technical leadership: architecture, roadmap and validation criteria.

1. DataGIS and operational grid data
2. Digital twinmodel of the low- and medium-voltage network
3. Load flowwhere lines and transformers hit their limits
4. Scenariosgrowth of PV, e-mobility and heat pumps
5. Plangrid expansion plan for regulatory reporting

Used by grid operators including

• Stuttgart Netze
• Netze BW
• Schleswig-Holstein Netze
• Netzgesellschaft Düsseldorf
• Netzgesellschaft Niederrhein

grydnova.com →

IP knowledge graph co-developed · live

A live knowledge graph of OmegaLambdaTec's intellectual property, generated directly from its code repositories.

• Traceable: every method links to the code, commits and assets behind it
• Connected: edges show how methods depend on and evolve from each other
• Scored: each method is rated by maturity (foundational, specialised, established) and by reuse across products and industries

Physics methods for industrial optimisation energy optimisation

1. From astrophysicsadaptive time-stepping: small steps only where the physics demands it
2. To industryenergy-management optimisation solvers
3. Resultup to 100–1000× faster convergence, reused across projects

Earlier at OmegaLambdaTec: a linear-programming energy-management system built from first principles with a German university, and a day-ahead gas-price forecast that placed on the podium in the Ruhr Challenge 2018.