☀️
Light

Sean Andrie Delfin Gadingan

ML / Software Engineer  ·  Software Development Student @ 42 Abu Dhabi
Abu Dhabi, UAE · seanandriegadingan@gmail.com · LinkedIn · GitHub · Portfolio
Experience
AI Engineer Intern — Al Hathboor Bikal.AI (AHB.AI)
Jul – Oct 2024
Sharjah, UAE
  • Contributed to an end-to-end machine learning pipeline for automated glaucoma detection from retinal fundus images, targeting early-stage screening in clinical environments with limited specialist access.
  • Fine-tuned a MobileNetV3-Large backbone in PyTorch on ~8,770 retinal fundus images from the EyePACS-Light v2 dataset, achieving 91% accuracy and 0.91 AUROC on the held-out test set for binary glaucoma classification.
  • Developed comprehensive data preprocessing, analysis, and augmentation pipelines — including random flips, contrast adjustments, and resolution normalisation — to improve model robustness against real-world image variability.
  • Designed and deployed a clinical-facing glaucoma detection application, integrating the trained model behind an inference API for real-time predictions from uploaded fundus images.
Education
42 Abu Dhabi
Dec 2024 – Present
Diploma in Software Engineering
Relevant: C/C++, Network Administration, Graphics Programming, Unix Systems
Bath Spa University
BSc Creative Computing — First Class Honours (GPA: 3.8)
Relevant: Machine Learning, Advanced Programming, Research Methods, Emerging Technologies
Hackathons
DataMind RecognAIze Hackathon — 2nd Place
Jun 2025
42 Abu Dhabi
  • Designed a face re-identification pipeline using InsightFace to extract 512-dimensional facial embeddings, then trained and benchmarked multiple Scikit-Learn classifiers to perform identity matching across varied lighting and pose conditions.
  • Enhanced raw embeddings by concatenating auxiliary signals — facial landmarks, estimated head pose, and detection confidence scores — to enrich identity representations and improve classifier discriminability.
  • Conducted systematic hyperparameter optimisation of Ridge Classifier and Linear SVM using Optuna across 10-fold stratified cross-validation, achieving a macro-F1 of 0.90 with lightweight linear classifiers that require no GPU at inference time.
  • Validated embedding quality and identity separability via UMAP projection, confirming per-identity clustering and diagnosing failure modes in confusable identities.
SambaNova Agentic AI Hackathon — 3rd Place
Jan 2025
42 Abu Dhabi
  • Developed Rihlat, a multi-agent AI assistant for Dubai public transport, powered by Llama 3.1-70B via the SambaNova API, enabling natural-language queries over complex real-world transit data.
  • Built a suite of 4 modular LangChain tools covering GTFS schedule querying, address geocoding via TomTom, turn-by-turn routing via Bing Maps, and temporal reasoning for timetable-aware journey planning.
  • Designed a dynamic prompt templating system generating well-structured SQL queries and structured API call payloads at runtime, enabling the agent to handle a wide range of transit queries without hardcoded logic.
  • Implemented zero-shot tool orchestration, allowing the agent to autonomously select and chain tools at inference time for flexible multi-step reasoning.
Projects
libtensr — N-Dimensional Tensor Library in C
2025
  • Designed a lightweight, dependency-free tensor library in pure C with stride-based row-major memory layout, automatic stride computation, and NumPy-style implicit broadcasting — targeting embedded systems and low-level ML inference engines.
  • Built a generic elementwise operation framework driven by user-defined callbacks, and developed multi-dimensional index iteration utilities for correct traversal of non-contiguous and transposed tensor views.
DocuMate — RAG Document Q&A Chatbot
2023–2024
  • Built a RAG chatbot supporting PDF, DOCX, and TXT uploads using FAISS vector search and OpenAI embeddings to ground answers in source material and reduce hallucination risk on domain-specific documents.
  • Implemented a history-aware retrieval architecture that reformulates follow-up questions against prior conversation context before retrieval, enabling coherent multi-turn Q&A; deployed on Streamlit Cloud with Firebase-backed session persistence.
Structure-from-Motion with LightGlue
2023–2024
  • Developed a Streamlit-based interactive tool for 3D point cloud reconstruction from unordered image datasets, integrating LightGlue with configurable feature matchers (SuperPoint, DISK) for keypoint extraction and cross-image correspondence matching.
  • Enabled GPU-accelerated reconstruction via Google Colab with automated tunnelled remote access, allowing users to offload compute-intensive triangulation steps without local GPU hardware.
BukoShell — Unix Shell in C
2025
  • Built a POSIX-compliant Bash-like shell in C from scratch with command parsing, environment variable expansion, pipelines, I/O redirections, heredocs, and process execution via fork/execve.
  • Added wildcard glob expansion, logical operators (&&, ||), and subshell execution, extending expressiveness to support common real-world scripting patterns.
FdF — 3D Wireframe Renderer
2024
  • Developed a real-time interactive 3D wireframe renderer in C using MiniLibX with isometric and perspective projection modes, affine transformation matrices for translation, rotation, and scaling, and dynamic colour gradients mapped to elevation values.
Skills
Languages Python, C, C++
Frameworks PyTorch, Scikit-Learn, TensorFlow, NumPy, Pandas, XGBoost, LangChain, Streamlit, Firebase, FAISS, SQL, Git
ML Neural Networks, Generative AI, Computer Vision, Transfer Learning, Classification & Regression, Data Analysis, Visualization