Cesilia Nainggolan

I'm still early in my career and keen to learn more. These are projects I built to practise, and each one says what it measured and where it falls short.

Streamlit app that answers road crash questions in plain English llm-project (crashsql) Answers plain-English questions about 202,554 Victorian road crashes. 82% correct on a 22-question gold set (measured).

Skills

Machine learning and computer vision

PyTorch, TensorFlow, scikit-learn, ONNX
Computer vision, NLP, LLM agents, recommenders
CNNs, transfer learning, model evaluation

Software engineering

Python, C/C++, TypeScript, JavaScript, Java, C#
FastAPI, Node.js, React Native, .NET
Embedded (STM32, ESP32), Docker, Terraform
Azure, GitHub Actions, Git, Scrum

Data analysis and visualisation

SQL, PostgreSQL, SQL Server, dbt, pandas
Power BI, Tableau, Excel

Emoji mirror demo page with a blurred webcam frame and probability bars facial-expression-lab Re-tested my thesis approach on JAFFE. Accuracy fell from 0.864 to 0.573 once the test set held out whole people (measured). Traffic volume demo page built on real October 2026 data mini-city-brain Replayed 1,577,675 real traffic readings through an API, queue and warehouse. 15 of 15 data checks passed (measured, one laptop run). Courier's Atlas puzzle game: pixel-art map with named districts cs-notes-and-games Free computer science study site and puzzle game: 104 levels, 68 chapters. Built with Claude Code, disclosed in the README. Indonesia wheat trade dashboard showing title and key figures asia-commodity-trade Australia's share of Indonesia's wheat imports fell from 57.6% to 40.6% between 2015 and 2025 (UN Comtrade). Confusion matrix and float32 against partly int8 model size and accuracy, measured on the test split edge-vision-lab Scene-split bird, drone and neither classifier. Plain int8 broke it; a partly int8 model recovered most of it (measured, laptop CPU, no board).

Get in touch

Happy to talk about graduate and junior roles.