M.Sc. student in Image Analysis and Machine Learning at Uppsala University (Sep 2025 – Jun 2027), building on a B.Sc. in Data Science from NOVA IMS (Lisbon) and an Erasmus semester at Lund University. I work on computer vision and deep learning, and I'm currently a research intern at Furhat Robotics (Stockholm) on multimodal perception. Portuguese, based between Stockholm and Uppsala.

Computer vision & medical imaging

• Semi-supervised distillation for multimodal cancer-cell classification — on a hard, patient-disjoint 12-patient Kaggle benchmark, I grew the effective training set with hard pseudo-labels (Lee 2013) then soft-target knowledge distillation (Hinton 2015) on a dual EfficientNet-B0 with AdaBN — the change that mattered more than scaling the network. Full LaTeX methodology write-up and an honest failure analysis. → multimodal-microscopy-distillation
• Vision Transformers for emotion recognition — fine-tuned ViT-Base on RAF-DB for facial-emotion recognition, feeding the dialogue layer of a Furhat social-robot tutor (group project). → furhat-emotion-tutor
• Classical image analysis from the toolbox up — hand-written histogram operations, Otsu and iterative thresholding, colour segmentation, spatial filtering, multispectral Landsat imagery, and rigid image registration (intensity- and feature-based) in MATLAB. → classical-image-analysis

Machine learning & statistics

• Time-to-detection in microfluidic bacterial growth — benchmarking Gaussian processes, changepoint detection, bootstrap, HMMs, and functional data analysis to find the earliest reliable call that an antibiotic is working. → bacterial-growth-detection
• Gradient-boosted pass-danger model — flags passes likely to create a shot from StatsBomb event data and ranks passers across five leagues, with pitch-control and circular-statistics run analysis. → football-pass-danger-model
• Deep learning from scratch — a NumPy-only fully-connected network with hand-written backprop, then a PyTorch progression to a residual CNN on MNIST. → mnist-from-scratch-to-resnet

LLMs, multimodal & human–robot interaction

• Furhat social robot for student mental-health check-ins — a group HRI study (1MD043) testing whether empathetic vs. neutral robot behaviour changes how students perceive a wellbeing check-in (between-subjects, n = 23, RoPE + Godspeed scales); I contributed to the experimental design and analysis and added a Gemini LLM backend. → HRI_Furhat
• Multimodal mushroom-ID chatbot — a Gradio app that reads a photo, extracts a strict-JSON description, and answers as a mycology assistant with a safety guard against foraging advice; runs on Google Gemini or a Hugging Face pipeline (Llama 3.1 + BLIP captioning). → mushroom-id-chatbot
• LoRA fine-tuning of Qwen3-0.6B — a small-LLM SFT experiment on a BeautifulSoup-scraped corpus, with a Gradio base-vs-fine-tuned comparison. → greek-mythology-chatbot

Languages: Python (daily), SQL, R, MATLAB, Bash Deep learning & computer vision: PyTorch, Hugging Face Transformers, timm; Vision Transformers, CNNs, transfer learning, knowledge distillation, semi-supervised learning, PEFT / LoRA Also: scikit-learn, scikit-image, NumPy, SciPy, Pandas; Gradio; Git, Jupyter, Kaggle; LLM APIs (OpenAI, Gemini)

Research intern at Furhat Robotics (Jun – Dec 2026). Looking for computer-vision, ML / AI engineering, and deep-learning research roles — and a master's thesis — in Sweden and the EU.

• Email: rafaeltproenca@gmail.com
• LinkedIn: linkedin.com/in/rafael-alexandre-proenca