Profile

👤 Summary

Research Engineer specialising in continual learning and LLM post-training. Currently at Huawei Paris Research Center working on how LLMs store and update knowledge, with focus on catastrophic forgetting and episodic memory in Transformers. My master thesis at Italy’s National Cancer Institute produced 4 peer-reviewed publications on CT-based survival prediction in lung cancer, across IEEE EMBC, ESMO AI, ECAI 2025, and ECML PKDD.

Along the way I won a 36-hour hackathon with an LLM mobile app (Aurora), led GenAI prototypes for 5+ enterprise clients at Salesforce, and spent two years as software engineer and then team leader for Policumbent, winning the IHPVA world human-powered speed record twice.

MSc, Politecnico di Milano 110/110  ·  BSc, Politecnico di Torino 110/110 cum laude Alta Scuola Politecnica  ·  Intraprendenti honor program

💼 Experience

Research Engineer

Huawei | Paris, France | 10/2025 – Present

Working on episodic memory and continual learning in LLMs, studying how models store and update knowledge without catastrophic forgetting.

• Designed and maintained training pipelines for 1,000+ multi-GPU full fine-tuning runs on Mistral and LLaMA models using DeepSpeed ZeRO-2.
• Built gradient-tracking infrastructure to regularize fine-tuning and reduce catastrophic forgetting.
• Built an end-to-end synthetic data pipeline generating 21K+ high-quality training samples; developed an LLM-as-judge evaluation system achieving 98–99% agreement with human annotators across 500+ test cases.

Graduate Student Researcher: Master Thesis

Fondazione IRCCS Istituto Nazionale dei Tumori di Milano | Milan, Italy | 09/2024 – 12/2025

My thesis was about predicting 6-month overall survival in advanced NSCLC patients from lung CT scans, with explainability and fairness checks built in, not just accuracy.

• Developed and evaluated 2D and 3D pipelines on the Apollo11 cohort (385 patients).
• Benchmarked CNN-based models (ResNet50) against Vision Transformers and GAN-based approaches.
• Applied SmoothGradCAM++ to identify which CT regions drove predictions.
• Showed that naive augmentation strategies can silently introduce demographic bias.
• Achieved 0.74 F1 on 6-month overall survival prediction, a clinically meaningful and technically hard task with significant class imbalance.
• 4 peer-reviewed publications from a single master thesis: IEEE EMBC 2025, ESMO AI 2025, AEQUITAS @ ECAI 2025, PharML @ ECML PKDD 2025.
• Reviewer for the 47th IEEE EMBC conference.

Student Researcher: Computer Vision and Robotics

PIC4SeR / Yanmar S.P.A. (Alta Scuola Politecnica) | Turin, Italy | 04/2024 – 09/2025

Yanmar wanted to automate grapevine pruning. The approach: detect vine structure from images, reconstruct the topology as a graph, then apply agronomic rules over it to decide which canes to cut.

• Built a keypoint detection pipeline on stacked Hourglass Networks for vine junction and cane localisation.
• Designed a graph-based reconstruction module converting 2D detections into vine topology.
• Integrated heuristic pruning strategies derived from expert agronomic knowledge.
• Achieved 95% precision, 90% recall, 92% F1 on the keypoint detection module.
• GitHub: AlbertoEusebio/VinPRO

Solution Engineer Intern

Salesforce | Milan, Italy | 09/2024 – 07/2025

Client-facing prototyping: building and demoing GenAI systems for enterprise clients, typically on short timelines.

• Built AI agents on the Salesforce Agentforce platform for autonomous task execution.
• Delivered proofs of concept to 5+ large businesses, adapting to different industries and integration constraints.
• Prototyped GenAI workflows for sales and service automation: case routing, response generation, pipeline management.
• Everything had to be deployable within real enterprise security constraints, not just a slides demo.

Graduate Student Researcher: NLP

NECSTLab, Politecnico di Milano | Milan, Italy | 11/2023 – 06/2024

• Built Aurora, a cross-platform mobile app using a fine-tuned Meta-Llama-3.1-70B (via AWS Bedrock) for daily mental well-being tracking and structured self-reflection.
• Users share short voice reflections; the app organises them across eight well-being dimensions and suggests activities.
• Designed the serverless backend on AWS Lambda; front end in React Native and Expo.
• Ran a 27-participant user study to evaluate the interaction model and response quality.
• Won Hack the NECSTCamp (36-hour AI mobile app challenge) 🏆.

Team Leader

Team Policumbent | Turin, Italy | 10/2022 – 10/2023

Led the engineering program for Team Policumbent: 9 divisions, 4 vehicles, ~100 students, 23 sponsors.

• Coordinated all technical activities across the team and represented Policumbent at Automation and Testing Turin and Maker Faire Rome.
• Won the IHPVA World Human Powered Speed Challenge 2023 and matched the world record 🏆.

Software Engineer

Team Policumbent | Turin, Italy | 03/2021 – 10/2022

First engineering role: embedded systems and telemetry for Cerberus and Phoenix, two high-speed human-powered streamliners.

• Built Raspberry Pi software interfacing with Garmin ANT+ sensors for real-time biometric data logging.
• Developed an ESP32-based weather station for race-condition monitoring.
• Implemented MQTT-based telemetry supporting vehicles running at 130 km/h.
• Improved rider performance by 20% on the Cerberus handbike workstream.
• Contributed to the IHPVA World Human Powered Speed Challenge win in 2022 🏆.

🎓 Education

📄 Publications & Research Output

🛠️ Skills

ML / AI

Trustworthy AI

Frameworks & Tools

Infrastructure

Languages & Systems

📬 Contact

• LinkedIn: linkedin.com/in/alberto-eusebio
• GitHub: github.com/AlbertoEusebio
• Google Scholar: scholar.google.com
• Email: albertoeusebio72@gmail.com
• Languages: Italian (native), English (C1, CAE), German (A2, Goethe)