AI Infrastructure · AI SRE & AIOps · MLOps · HPC
Research Fellow, University of Bologna · Bologna, Italy
I build autonomous AI systems that act on infrastructure — not just explain it.
I came to this the unusual way round. Before the PhD I spent seven years as the IT and network administrator of a combined-cycle power plant of more than 1,000 MW: no staging environment, and a bad change measured in megawatts rather than in error budgets. Then a doctorate in high-performance computing at the University of Bologna, and since then research and platform engineering on EuroHPC-funded projects.
That background is why my systems ask before they act. Everything below is one attempt at the same question: how much of an operator's judgement can a machine take over, and how do you prove afterwards that it was right to?
📍 Bologna, Italy · remote across the EU · 🧭 mskazemi.com · 💼 available for freelance work
KubeIntellect — a human-governed AI SRE for Kubernetes
Ask your Kubernetes cluster a question in plain English. It gathers live evidence from tools such as kubectl, Prometheus and Loki, works out what is actually wrong — and pauses for your explicit approval before it changes anything.
- It executes cluster operations, behind a human approval gate — that is the whole point, and it is what separates it from a chat window next to a terminal
- LangGraph FSM supervisor with PostgreSQL checkpoints and human-in-the-loop approval gates
- Dynamic Code-Generator agent: writes and validates new tools at runtime, with registration routed through human approval
- Modular domain agents: logs, metrics, RBAC, lifecycle, scheduling, exec, proxy
- Peer-reviewed — Journal of Grid Computing (2026), 10.1007/s10723-026-09837-6
YazSes — offline voice dictation that never phones home
Hold a key, speak, release. Speech-to-text runs entirely on your own CPU and the words are typed into whatever window has focus — editor, browser, terminal, a remote SSH session. No cloud, no account, no subscription.
- Works on Wayland, where most dictation tools silently fail — injection happens at the kernel uinput layer, below the compositor, so GNOME, KDE, sway and Hyprland all behave the same
- Accuracy measured, not asserted: 2.59% WER on LibriSpeech test-clean (
small.en), 4.07% with the defaultbase.en— method and repro commands published - Also transcribes recordings and captures whole meetings with speaker labels, same install
- Built in the open by outside contributors — good first issues are tagged and waiting
| Project | What it does | Evidence | Stack |
|---|---|---|---|
| NovaFabric | A time machine for AI systems — capture, replay, diff and audit any agent or HPC run, with no code changes. Self-hosted. | Apache-2.0 · experimental · novafabric.ai | Python, OpenTelemetry |
| AOBench | Benchmark for LLM agents doing real HPC operations work. Role-aware and permission-enforced: a policy violation hard-fails the task, however correct the answer looked. | 88 tasks (10 categories × 5 roles) · 29 environment bundles, 6 from real Marconi100 telemetry · archived with a DOI · paper under review | Python, MCP, Slurm |
| ExaMLOps | End-to-end MLOps platform for HPC, built for the EuroHPC SEANERGYS project. Any of the sixteen consortium partners registers a model; the platform trains, versions, governs and serves it. Architect and lead developer. | In production at LuxProvide (MeluXina) · train → version → govern → serve, behind a sysadmin approval gate | Prefect, MLflow, Ray Serve, Slurm, FastAPI, React |
| kube-q | CLI and Python SDK for KubeIntellect — pip install kube-q |
Streaming responses, Rich TUI · AGPL-3.0 | Python |
| GRAAFE | Graph neural network that anticipates compute-node anomalies on exascale HPC, trained offline and served online through a Kubeflow pipeline on live telemetry. | Published, FGCS 2024 · CINECA Marconi100 | Python, GNN, Kubeflow |
| HazardNet | Thermal-hazard prediction for datacenters, over a year of temperature and power telemetry from 3,312 nodes of CINECA's Marconi A2. Six-hour horizon, chosen with the facility manager. | Published, FGCS 2024 · 1 GB dataset on Zenodo, CC BY 4.0 | Python, TCN/LSTM/SVM |
Remote across the EU, in three areas — each starting with a fixed-price audit so you can see the work before committing to a project:
- Kubernetes reliability & AIOps — health-checks, observability, hardening, incident root-cause analysis
- MLOps & ML in production — registry, serving, drift detection, governed retraining, monitoring
- Production LLM agents — tool boundaries, scoped RBAC, human-in-the-loop safety, tracing, audit trails
Based in Bologna, Italy — working remotely across the EU on CET.
→ Engagements, rates and how it works
PhD: Design, Analysis, and Management of High-Performance Computing Systems · University of Bologna (2018–2022)
EU projects: SEANERGYS (EuroHPC-JU — MLOps platform architect & lead developer, WP task lead) · DECICE · Graph-Massivizer · EUROPEAN PILOT · REGALE · EPI SGA1
Impact (Google Scholar, read 2026-08-10 — see the live profile for current figures): 218 citations · h-index 8 · i10-index 7, across 17 published or accepted peer-reviewed works. The trend is the interesting part: 9 citations in 2022, 15 in 2023, 24 in 2024, 76 in 2025, and 65 in the first seven months of 2026 — more in the last twenty months than in the preceding eighteen years combined.
| Title | Venue | Year |
|---|---|---|
| KubeIntellect: A Modular LLM-Orchestrated Agent Framework for Kubernetes Management | Journal of Grid Computing | 2026 |
| M100 ExaData: A Data Collection Campaign on CINECA's Marconi100 Tier-0 Supercomputer | Nature Scientific Data | 2023 |
| PM100: A Job Power Consumption Dataset of a Large-Scale Production HPC System | SC'23 Workshops | 2023 |
| GRAAFE: GRaph Anomaly Anticipation Framework for Exascale HPC Systems | FGCS | 2024 |
| HazardNet: A Thermal Hazard Prediction Framework for Datacenters | FGCS | 2024 |
| Elevating Datacenter Resilience with ThermADNet: A Thermal Anomaly Detection System | FGCS | 2026 |
| Multi-level Anomaly Prediction in Tier-0 Datacenter | ACM Computing Frontiers | 2022 |
Open data. I co-author three CC BY 4.0 datasets from CINECA's Tier-0 supercomputers — roughly 26 GB in total: M100 ExaData (24.8 GB of Marconi100 telemetry), the HazardNet thermal dataset (3,312 nodes, first author) and PM100 (per-job power). Free to download, commercial use permitted.
Full list, with current citation counts → Google Scholar · ORCID · dblp · OpenAlex
| Area | Tools |
|---|---|
| Platform & infrastructure | Kubernetes · Helm · Terraform · Docker · Linux · Azure |
| AI / ML | Python · PyTorch · LangGraph · FastAPI · MLflow · Ray Serve |
| HPC | Slurm · MPI · OpenMP · Prefect |
| Observability | Prometheus · Grafana · Loki · OpenTelemetry |
| Data & ML systems | GNNs · TCN/LSTM · anomaly detection · time-series telemetry at datacenter scale |
PC Member: PDP 2025 · PDP 2026 · AsHES 2026
Reviewer: IEEE TCAD · FGCS · Journal of Grid Computing · SC · ACM CF · DATE · PDP · AsHES
Supervision: 2 PhD co-advisees (ongoing) · 5 MSc theses completed · Lab of Big Data Architectures, UniBo (2020–2024)
Hiring, or need an audit? → mskazemi.com/hire
Website ·
About ·
LinkedIn ·
GitLab ·
Mastodon ·
Scholar ·
ORCID