Marcos Costa

Machine Learning Engineer

Machine Learning  ·  Medical AI  ·  Robotics  ·  Technology

I build software at the intersection of intelligence and the physical world — with a focus on machine learning, medical AI, and tools that make a real difference. My background as a professional climber for over 12 years taught me to push limits and think creatively under pressure; I bring that same drive to every engineering challenge.

Marcos Costa

Experience

Jun 2023 – Present

Software Engineer

Universitätsklinik Balgrist / ETH Research Group, Spine Biomechanics — Zürich

Developed and optimised client-server communication pipelines, increasing CT scan analysis efficiency by 30%. Enhanced medical imaging processes by developing machine learning algorithms for automatic segmentation of CT scans. Built full-stack solutions leveraging Python, Kotlin, Docker, AWS, REST APIs, and SQL databases.

May 2022 – Sept 2022

Machine Learning Engineer Intern

Universitätsklinik Balgrist / ETH Research Group, Spine Biomechanics — Zürich

Supported the development and enhancement of a CNN-based pipeline for automated segmentation of spinal CT imagery. Managed the end-to-end ML workflow from data acquisition and preprocessing to model training, evaluation, and performance tuning. Employed Python, SimpleITK, Slicer, and MedPy for robust, accurate model predictions.

Feb 2015 – Mar 2022

Professional Athlete

Global

Professional climber and alpinist with expertise in rock, ice, and high-altitude ascents. Successfully summited challenging global peaks including Gasherbrum II (8,035m). Twice nominated for Brazil's prestigious Golden Carabiner Award for Best Alpinist. Organised logistics for extensive expeditions across the Himalayas and Patagonia, and secured sponsorships from prominent global brands.

Jun 2013 – Feb 2015

Founder & COO

Vertical Horizon — Shanghai, China

Founded and operated an innovative educational enterprise blending outdoor adventure with impactful teaching. Designed outdoor programmes that measurably improved teamwork and problem-solving, and integrated community-driven humanitarian projects to enhance student engagement in local communities.

Projects

Deep Learning for Metal Artifact Reduction in CT Imaging

Deep Learning for Metal Artifact Reduction in CT Imaging

Bachelor thesis investigating deep learning architectures for metal artifact reduction (MAR) in CT scans. Compared dual-domain networks (InDuDoNet, InDuDoNet+) and image-domain methods (DICDNet) against state-of-the-art baselines. Explored supervised, unsupervised, and semi-supervised approaches to reduce streak artifacts while preserving anatomical detail.

Deep Learning Computer Vision Medical Imaging CT Imaging PyTorch
Real-Time Chinese Hand Gesture Classifier

Real-Time Chinese Hand Sign Classifier

A Streamlit app to build your own Chinese number hand-sign classifier (数字 0–9) from scratch. Collect training data via webcam, train an MLP on MediaPipe hand landmarks, and run live inference — all in the same browser UI.

Python Computer Vision MediaPipe scikit-learn Streamlit
S G

2D Robot Path Planning Visualizer

Interactive browser demo visualising five path planning algorithms — A*, Dijkstra, BFS, Greedy Best-First, and RRT — on a drawable obstacle grid. Watch each algorithm explore in real-time, compare stats, and adjust speed on the fly.

JavaScript A* RRT Algorithms Robotics
Swiss Avalanche Risk Dashboard

Swiss Avalanche Risk Dashboard

Live interactive dashboard scraping real accident data from WSL SLF — Switzerland's official avalanche authority. Four map views (scatter, heatmap, hexagon grid, 3D columns), analytics charts, and live webcam feeds for Andermatt, Engelberg, and Klewenalp.

Python Streamlit pydeck Plotly Web Scraping
Car Racing DQN Variants

Car Racing — DQN Variants

Applied three progressively improved Deep Q-Network algorithms (DQN, Double DQN, Dueling Double DQN) to OpenAI Gymnasium's CarRacing-v3 environment. Each variant builds on the last — fixing Q-value overestimation and separating state value from action advantage — with the Dueling Double DQN achieving the highest and most stable rewards.

Python Reinforcement Learning DQN PyTorch OpenAI Gymnasium
Flappy Bird DQN

Flappy Bird — DQN vs Double DQN

Trained a Deep Q-Network agent to play Flappy Bird, documenting a real training plateau caused by Q-value overestimation in vanilla DQN. Double DQN resolves it by decoupling action selection from evaluation, pushing past the score-62 ceiling and achieving significantly higher performance.

Python Reinforcement Learning DQN PyTorch TensorBoard

Awards

🏆

1st Place — Base Network AI Hackathon (Coinbase)

London, UK

Feb 2025

Built Base Super Assistant, an AI-powered trading assistant that automates on-chain trading through natural language processing and ML-driven decision-making. Leveraged OpenAI's API and LangChain for voice and text-based trade execution via Telegram, integrated FastAPI and the Base SDK for real-time blockchain transactions, and applied AI-based risk assessment for automated trade execution.

🏆

1st Place (Team) — AI/ML Trading Hackathon (EPFL)

Jump Trading × EPFL Trading Blitz — Lausanne, Switzerland

Oct 2025

Won 1st place as a team (4th individually) at a high-pressure quantitative trading hackathon with 75 participants and live, rapidly shifting market simulations. Collaborated in a cross-functional team to design and execute disciplined trading strategies under macro-driven volatility over a six-hour session.

Education

MSc Artificial Intelligence

ZHAW — Zürich University of Applied Sciences

Dec 2026

MS Computer Science

University of Colorado Boulder

July 2027

BSc International IT Management & Computer Science

Lucerne University of Applied Sciences and Arts (HSLU)

Sept 2022 – Jul 2025

Thesis: Effectiveness of deep learning architectures for metal artifact reduction (MAR) in CT imaging. Supervisor: Javier Montoya.

Micro Masters in Data Science

University of California San Diego (UC San Diego / edX)

Feb 2023 – Feb 2024

Business Administration

Loyola Marymount University — Los Angeles

Sept 2002 – Sept 2004

Beyond the Code

Marcos Costa — Alpinist

Professional Climber & Alpinist

Long before writing my first line of code, I was climbing mountains. As a professional alpinist for over 12 years I've led expeditions across the Himalayas and Patagonia, summited Gasherbrum II (8,035m), and been twice nominated for Brazil's Golden Carabiner Award for Best Alpinist. The same focus and resilience I developed on high-altitude walls is what I bring to every engineering challenge.

I'm also the founder of MNT Token — a blockchain project born from this passion, building a Web3 community around outdoor adventure and exploration.

Contact

Interested in collaborating, discussing a project, or just want to say hello? Feel free to reach out — I'm always happy to connect.