Open to ML engineering roles · Turin, IT

Large language models, built to run on the device in your hand.

Eziyo EhsaniMachine Learning Engineer

Quantization, distillation and retrieval at the edge — working out what a language model actually costs in latency, energy and accuracy once it has to run privately, without the cloud.

Explore the work Read the paper

Find me

S2 Group · Amsterdam

0LLMs on Snapdragon 8 Elite

0Isolated inference runs

Research & work with

S2 Group, Amsterdam University of Naples Federico II DIETI Lab HascoNet Preprint on arXiv Anthropic certified Empirical Software Engineering · Under review

01 — About

Systems that run where the data already is.

Based in Turin, Italy
MSc Data Science · Federico II · 108/110
AI Systems Engineer Intern · S2 Group
Manuscript under review · EMSE

I am a Data Science MSc graduate from the University of Naples Federico II (108/110) specializing in on-device large language models, model compression, and agentic AI systems, with a BSc in Computer Engineering from Azad University and 5 years of professional experience across data science and ML engineering.

My work builds intelligent systems that run where the data already is, from deploying eight open-source LLMs (0.5B–9B) on a Snapdragon 8 Elite handset with llama.cpp to profiling 480 isolated inference runs for latency, energy and accuracy. My MSc thesis at Federico II quantified the energy–accuracy trade-offs of on-device intelligence, and I currently work as an AI Systems Engineer Intern at S2 Group in Amsterdam. I am a fast learner who ramps quickly into new technical domains. I have a manuscript under review at Empirical Software Engineering, with a preprint available on arXiv.

energy → accuracy →

Thesis · Green Mobile Intelligence

Sustainability is not linear.

Quantifying energy–accuracy trade-offs in on-device LLMs across eight models, four quantization schemes and real mobile silicon.

On-device

0LLMs

0.5B–9B parameters, deployed with llama.cpp

Profiled

Isolated inference runs: latency, energy, accuracy

Modeled

0k⁺

Records behind churn & HR prediction models

MSc grade

0/110

Data Science, University of Naples Federico II

Experience

0yrs

Across data science and ML engineering

Languages

English C1/C2 Persian Native Italian B1

Under review

Empirical Software Engineering

arXiv:2603.26603

02 — Experience & Education

From dashboards to on-device inference.

Five years across data science and ML engineering, now focused on green AI and language models that run under real memory, compute and energy budgets.

Apr 2025 — Mar 2026Amsterdam, NL

Current role Apr 2025 – Mar 2026

AI Systems Engineer Intern

Software and Sustainability (S2) Group, Amsterdam

Green AI & On-Device Inference

Deployed and optimized 8 open-source LLMs (0.5B–9B) on-device using llama.cpp, PyTorch, and compression techniques (INT4/INT8 quantization, pruning, distillation), enabling agent-ready NLP inference under strict memory and compute constraints
Built automated Linux/Bash orchestration pipelines for 480 evaluation runs, collecting ML metrics (latency, energy usage, token throughput, accuracy) to validate autonomous, privacy-preserving language model deployments
Profiled latency–energy trade-offs across ARM-based mobile hardware and SoC architectures, defining configurations that balance real-time NLP responsiveness with sustainable resource consumption

llama.cppPyTorchQuantizationDistillation Linux / BashEnergy ProfilingSoC Architectures

Sep 2025 — Feb 2026Naples, IT

Sep 2025 – Feb 2026

On-Device ML Intern

DIETI Lab — University of Naples Federico II

Research lab Naples, Italy

Trained and evaluated transformer-based NLP models (BERT, DistilBERT) for text classification, comparing full-precision baselines against quantized on-device variants. Applied explainable AI techniques (SHAP, LIME) to clarify model behaviour and support accuracy–efficiency trade-off decisions, and built supervised learning pipelines in Python (scikit-learn, PyTorch) balancing latency, energy use and predictive performance.

Jul 2020 — Aug 2023Tehran, IR

Jul 2020 – Aug 2023

Data Scientist

HascoNet

Analytics & ML Tehran, Iran

Automated KPI dashboards in Power BI, cutting manual reporting time by 30%. Built predictive churn and HR analysis models over 100k+ records to inform retention and staffing decisions, and designed the ETL pipelines feeding analytics and machine learning workloads.

Jan 2020 — Jun 2020Tehran, IR

Jan 2020 – Jun 2020

Data Science Intern

HascoNet

Business intelligence Tehran, Iran

Built Power BI data models and DAX measures enabling dynamic filtering and performance monitoring, and supported senior analysts with reporting, data validation and documentation across analytics projects.

Sep 2023 — Mar 2026Naples, IT

Education Sep 2023 – Mar 2026

MSc in Data Science — 108/110

University of Naples Federico II

Master's degree Naples, Italy

Thesis: Green Mobile Intelligence — Quantifying Energy-Accuracy Trade-offs in On-Device LLMs. Coursework in AI Systems Engineering, Data Mining & Machine Learning, and Hardware & Software for Big Data.

Sep 2017 — Jul 2021Tehran, IR

Education Sep 2017 – Jul 2021

BSc in Computer Engineering

Azad University

Bachelor's degree Tehran, Iran

Thesis: Design of an IoT-Based Energy Consumption Monitoring and Optimization System. Coursework in System Design & Analysis, Compiler Design, Microprocessors and Computer Architecture.

MSc ThesisFederico II

MSc Thesis

Green Mobile Intelligence.

Quantifying energy–accuracy trade-offs in on-device LLMs.

How far compression can be pushed before quality collapses, measured across eight open-source models, four quantization schemes and 480 isolated inference runs on real mobile silicon.

University of Naples Federico II · Naples, Italy

108/110Final grade

2023–26Sep 2023 – Mar 2026

03 — Selected work

Built end to end, measured on real hardware.

Systems built end to end — from quantized inference on real handsets to fully offline retrieval and speech.

On-Device LLM Optimization

Deployed 8 open-source LLMs (0.5B–9B) on a Snapdragon 8 Elite device using llama.cpp with Q4_K_M and IQ4_XS quantization; automated 480 isolated inference runs with BatteryManager energy profiling to quantify the feasibility of on-device agent deployments under real hardware constraints.

llama.cppQ4_K_M · IQ4_XSSnapdragon 8 EliteBatteryManager

View on GitHub

RAG Clinic Assistant

A tool-using RAG agent for a medical clinic (Python, LlamaIndex, ChromaDB, local LLMs) that retrieves grounded context and orchestrates multi-step query resolution, deployed via a FastAPI backend with a custom chat interface.

LlamaIndexChromaDBFastAPI

View on GitHub

Device Power Profiler

Extended an Android/Kotlin energy profiling tool interfacing with Android's BatteryManager API via ADB, adding reliable telemetry for on-device agent evaluation across heterogeneous hardware.

KotlinBatteryManager APIADB

View on GitHub

Local Voice Cloning & TTS

A fully offline multilingual voice cloning system (OpenVoice, MeloTTS, Gradio) enabling autonomous on-device voice synthesis in 6 languages with no cloud dependency.

OpenVoiceMeloTTSGradio6 languages

View on GitHub

04 — Research

Peer-reviewed evidence, not claims.

One manuscript under review, with the preprint openly available on arXiv.

2026

Under review

Empirical Software EngineeringEMSE · Springer

Sustainability Is Not Linear: Quantifying Performance, Energy, and Privacy Trade-offs in On-Device Intelligence

Eziyo Ehsani, Luca Giamattei, Ivano Malavolta, Roberto Pietrantuono (2026)

Read on arXiv arXiv:2603.26603

05 — Skills & Certifications

The toolkit, from model to metal.

Methods, frameworks and tooling used across research, deployment and production data work.

Core Methods

Foundations

Supervised LearningUnsupervised Learning Sequence ModelingReinforcement Learning Bayesian OptimizationCross-Validation Robustness Evaluation

Machine Learning & NLP

Modeling

PyTorchTensorFlowXGBoost BERT / TransformersWord Embeddings RAGspaCyBERTScore LLM-as-a-Judge

Agentic AI & Orchestration

Agents

LangChain / LangGraphCrewAIAutoGen Multi-Agent SystemsTool Use Function CallingPrompt Engineering n8nMCP

On-Device & Edge AI

Deployment

llama.cppINT4 / INT8GGUF AWQKnowledge DistillationPruning ONNX RuntimeSoC Architectures Latency & Energy Profiling

Programming

Languages & tools

PythonC++SQL Bash / ShellADBGDB Linux perf

Engineering & DevOps

Infrastructure

FastAPIDockerGit / GitHub CMakeCI/CDChromaDB SparkHadoopETL Redis

Certifications

4 verified

Building with the Claude APIAnthropic 2026 Model Context Protocol: Advanced TopicsAnthropic 2026 Claude Code 101Anthropic 2026 Data Analyst TrackDataCamp 2025