Federico Milana

2025 - under development

• Developing a Python library to estimate the training resource costs (time, money, and energy) for ML models
• Provides a web interface and a VS Code extension for easy integration into ML workflows
• Extracts and analyzes training parameters (e.g., model size, dataset size, model hyperparameters) from Python scripts
• Integrates LLM APIs to provide realistic estimates for different model architectures and training methodologies
• Designed to assist researchers and developers in planning and budgeting for ML training initiatives, optimizing resource allocation, and evaluating the environmental impact of their models

2025 - under development

• Fine-tuned LLaMA-3-8B-Instruct on personal WhatsApp group chats using Quantized Low-Rank Adaptation (QLoRA) for efficient training on consumer hardware
• Implemented parameter-efficient fine-tuning techniques to maintain model performance while reducing computational requirements
• Engineering a communication pipeline using the Telegram API for seamless integration, allowing the model to analyze conversations and generate contextually appropriate responses that mimic a specific user's writing style

2025

• An Electron application developed to visualize and compare manual and AI-generated annotations of English translations of "Conversations on the Plurality of Worlds" by Bernard Le Bovier de Fontenelle, 1686
• Extracted, aligned and segmented text to annotate with OpenAI's, Anthropic's and Gemini's APIs
• Developed for the forthcoming publication:
• Anna Maria Cipriani, Federico Milana (2025, Forthcoming). AI-Powered Corpus Translation Studies

2023 - 2024

• Trained XGBoost and fine-tuned BERT interpretable text classification models using LIME, SHAP, Occlusion values, Integrated Gradients and experimental LLM summarization of confusion matrices
• Conducted a user study with 128 participants to evaluate current model interpretability techniques

2021 - 2024

• An Interactive Machine Learning (IML) tool for thematic analysis of qualitative data, enabling iterative XGBoost model training and customization
• Conducted a user study with 20 participants to evaluate interaction with IML systems
• Developed for the following publication:
• Federico Milana, Enrico Costanza, Mirco Musolesi and Amid Ayobi (2025). Understanding Interaction with Machine Learning through a Thematic Analysis Coding Assistant: A User Study

2019

• A simulated social trading platform for investment decisions with chatbot guidance
• Conducted a user study with 64 participants to evaluate effects of UI and response variations on user trust
• Developed for the following publication:
• Federico Milana, Enrico Costanza and Joel E Fischer (2023). Chatbots as Advisers: the Effects of Response Variability and Reply Suggestion Buttons

2021

• Goal: Prototype explainability features in TACA (heatmaps, rationales, confidence cues) and design a study flow to assess their effect on qualitative coding
• Research: Reviewed XAI patterns (saliency, rationales, confidence) + heuristic walkthroughs with qualitative researchers
• Requirements: Toggleable explanations; preserve reading flow; show model limits; enable quick re-label with rationale
• Design: Desktop wireframes for Text (highlighted spans, inline rationales) and Keywords (theme buckets, re-classify) guided by usability heuristics
• Key UI: “Explain” toggle; tooltips; side theme bars; FP/TP galleries; confidence on demand
• Testing: Planned between-subjects study (with/without explanations) measuring accuracy, time, trust, and think-aloud insights
• Outcome: Clickable prototype + study protocol; findings guided later TACA iterations (final evaluation done separately)

2019

• Goal: Design an ICU dashboard for the University College Hospital to display real-time patient metrics, safety indicators, and unit performance
• Research: Interviews and contextual observations with ICU nurses to understand workflows, priorities, and pain points
• Requirements: At-a-glance patient status; clear safety alerts; ward-level performance view; minimal clicks to key metrics
• Design: Modular card-based layout for “Unit at a glance,” target metrics, safety metrics, and quality outcomes; colour-coded indicators for rapid scanning
• Key UI: Bed occupancy and usage panels; interactive ward map; metric cards with progress rings; split North/Southside performance; delay reasons colour-coded by type
• Testing: Iterative feedback sessions with ICU nurses to refine data grouping, colour use, and alert thresholds
• Outcome: Delivered high-fidelity dashboard prototype aligning with clinical workflows and supporting rapid decision-making

2018

• Goal: Increase community engagement with the Calthorpe Project, a community garden in central London
• Research: Interviews with staff and visitors to gather requirements; observed on-site interactions
• Concept: Initial vegetable vending machine and interactive menu ideas replaced with an outdoor interactive map to promote activities and volunteering
• Design: Low-fidelity prototype using cardboard, printed map, and tablet; map and touchscreen UI designed to guide visitors to points of interest
• Testing: Observed passers-by interacting with the prototype; noted usability issues and refined prompts and navigation
• Outcome: Validated map concept as effective for attracting attention and informing visitors about the garden

2017

• Goal: Design and evaluate a desktop airline booking experience
• Research: Competitive analysis (easyJet, BA, Ryanair) + 12-person questionnaire
• Requirements: Prioritise flight booking, allow booking without login, minimise pop-ups, highlight seat selection and extra baggage
• Design: 15 linked wireframes in Balsamiq guided by Nielsen/Molich principles and user mental models
• Key UI: Central flight search; top nav for check-in/cancel; step progress indicators; horizontal results to reduce scrolling; error-prevention via disabled actions
• Testing: 5 participants in natural settings; SUS survey + qualitative notes
• Findings: Improve visibility for “Home”/checkboxes; positive feedback on minimal look and navigation shortcuts
• Outcome: SUS 84/100 (above-average usability)

AI-Powered Corpus Translation Studies

Anna Maria Cipriani and Federico Milana

Springer, Forthcoming

• Developed a tool for visualizing and comparing manual and AI-generated annotations of English translations of "Conversations on the Plurality of Worlds"
• Analyzed translations using OpenAI's, Anthropic's and Gemini's APIs
• Demonstrated the potential of AI in corpus translation studies

Federico Milana, Enrico Costanza, Mirco Musolesi and Amid Ayobi

ACM SIGCHI Conference on Computer-Supported Cooperative Work & Social Computing (CSCW)

• Conducted a user study with 20 participants exploring how non-experts interact with Interactive Machine Learning (IML) systems
• Developed TACA, an IML tool for thematic analysis, enabling iterative model training and customization
• Found that users gained new analytical insights and adapted their interpretative approaches through IML interaction
• Identified key challenges in user understanding of ML concepts and proposed design recommendations for IML systems

Federico Milana, Enrico Costanza and Joel E Fischer

ACM conference on Conversational User Interfaces (CUI)

• Investigated how chatbot design features influence users' likelihood to follow AI advice in decision-making contexts
• Designed and conducted an incentivized study examining effects of response variability and reply suggestion buttons
• Built a simulated social trading platform where participants made investment decisions with chatbot guidance
• Found that both response variability and reply suggestions significantly increased users' trust and adherence to chatbot advice