Duarte Pereira da Cruz

• Joining The Tech Collective, Implement’s data and analytics unit, alongside the MSc.

• Co-designed the off-cluster gRPC control plane (Go) for an internal fuzz-testing platform, with two-way Kubernetes reconciliation, a TypeScript provisioning shim on Durable Objects, and a GitLab CI component other teams use. 40+ merge requests in three months.
• Found and fixed a concurrency race in the OAuth token refresh of an AI-agent gateway used internally and by customers. A SQLite-backed single-flight leader election removed ~9,700 errors a week, 97% of its failures.
• Hardened platform APIs against race conditions and cross-tenant leaks, backed by 490+ unit tests.

• Co-led an 8-person team to run a 24-hour AI hackathon for 150+ participants with €25,000+ in prizes; pitched sponsors and API partners, including NOS, Jane Street and ElevenLabs.

• Ran lab sessions for the Human-Computer Interaction course, a class of 24.

• Turned the team’s architecture conventions into Markdown and YAML rule sets that let LLMs generate complete Data, Core, IoC and API projects.
• Automated project scaffolding, cutting setup time by 87%.

Coursework: Perception for Autonomous Systems, Deep Learning, Deep Learning in Computer Vision, Robotics, Machine Learning Operations.

GPA 16/20. Academic Merit award, 2024/25. Grades: Machine Learning 20/20, Human-Computer Interaction 19/20, Probability and Statistics 18/20.

• Trained five classifiers on 63,000 European Portuguese articles; an MLP on TF-IDF reached 0.926 test F1 (BERT: 0.948).
• Explained them with LIME, L1 sparsity and topic clustering, and co-authored the paper. The model runs in the browser on this site.

• Built a segmentation dataset from Portugal’s COS2018 land-use map and fine-tuned YOLOv8n-seg to outline invasive acacia in satellite imagery.
• Wrote the FastAPI, Celery and Redis backend for asynchronous inference on uploaded imagery.

• Built the backend for a meeting assistant on Snapchat Spectacles: WebSocket audio streaming, silence-based chunking for low-latency captions, LLM summaries and task extraction, and text-to-speech.

• Built most of the FastAPI backend, the Gemini TTS and video pipeline and most of the React frontend for an agent that turns your feeds into a personal audio briefing; deployed it with Docker on Railway.