EPFL → ETH · Applied ML · Builder desk
I work on applied machine-learning systems where the data pipeline matters as much as the model. Recent work spans medical imaging research, recommendation and credit-risk systems, retrieval/RAG products, fraud-detection tooling, and technical project delivery.
Researcher with builder instincts.
I like technical work that starts with messy inputs and ends in a system someone can actually use.
Current thread
Medical imaging, recommendation systems, retrieval/RAG, and practical ML delivery.
CT angiography, vessel signal, segmentation
Leakage checks, geometry, experiment loops
Models that survive real constraints
Useful answers from messy documents
Next
ETH Zürich
MSc Robotics, Systems and Control
01Research
Idiap
Medical imaging + vascular ML · AUROC 0.840
02Delivery
Junior Entreprise
CHF 50K+ technical mandate portfolio
03Builder
MNT-Halan / EHL / EPFL IP
Recsys, RAG products, fraud-detection tooling
04Selected work
A few threads that connect the research, startup, and client-delivery sides of my work.
Industry
MNT-Halan
Built a content-based recommendation engine combining customer affinity and product popularity, engineered a 63-feature credit-risk dataset from 300k+ loan applications, and shipped a hybrid RAG pipeline (Dense + BM25 + Cross-Encoder) at 88% Acc@1 and 100% Acc@4.
Signal
Applied ML under production data constraints
Research
Idiap Research Institute
Developed a 3D CNN pipeline for renal-artery abnormality detection in CTA scans, rebuilt preprocessing to eliminate patient leakage and preserve anatomical geometry, and evaluated multiple image representations to a test AUROC of 0.840.
Signal
Preprocessing as part of the science
Builder
EHL Innovation Hub
Built the backend for a hotel-document RAG assistant, covering ingestion, chunking, embeddings, retrieval, response generation, and guardrails.
Signal
Useful AI from messy real documents
Operator
Junior Entreprise EPFL
Managed ML, AI, and software projects for clients while supporting recruitment, internal operations, student-association partnerships, and operating resources.
Signal
Technical judgment under delivery constraints
Experience log
The common thread is turning uncertain, messy technical work into something testable and useful.
MNT-Halan (Egypt’s First Fintech Unicorn)
Idiap Research Institute
Junior Entreprise EPFL
EPFL
EHL Innovation Hub
EPFL Innovation Park
Microsoft
MSc in Robotics, Systems and Control
Incoming master's program focused on robotics, systems, and control.
Bachelor's Degree, Electrical and Electronic Engineering
Coursework across Analysis I–IV, matrix & numerical analysis, probability & statistics, machine learning & data science, signal processing, control systems, embedded systems, telecommunications, and microwave & antennas.
Physics and Mathematics Summer Programme
Physics and Math course using Python libraries to simulate various phenomena.
Smaller builds and experiments that show the range: computer vision, voice systems, simulation, embedded programming, and ML baselines.
Implemented a complete FM stereo receiver in GNU Radio with demodulation, 19 kHz pilot recovery, stereo reconstruction, filtering, and 250-to-48 kHz resampling; validated live reception and an arctangent-free FM demodulator.
Built an end-to-end beverage classifier by scraping ≈2,000 images, fine-tuning ResNet18 with augmentations, and implementing a modular training pipeline with checkpointing.
Multilingual voice assistant supporting transcription, translation, email drafting, custom Arabizi normalization, and real-time browser recording.
Built a real-time homebrew game on ARM9/ARM7 processors, implementing timers, interrupts, graphics, audio, touchscreen input, and sprite management directly on the hardware.
Visual simulator of species population dynamics under constraints, with random algal growth and scavengers consuming algae.
Built a binary classifier to detect a virus from patient data.
Publications & thesis
Thesis and research writing connected to the medical-imaging thread.
Bachelor Thesis · EPFL & Idiap Research Institute · 2026
Working notes
Note 01I like work where the preprocessing is part of the science, not a cleanup step.
Note 02I look for failure modes early: leakage, distorted geometry, brittle retrieval, unclear assumptions.
Note 03I build in layers: data pipeline, baseline, experiment loop, then interface.
Note 04I'm most interested in applied ML that ends up being used by people.
Grouped by how I actually use them: research, product systems, foundations, and delivery.