Tu Phan · Machine learning · Software · Platform
Six years across three kinds of work: machine learning that runs in production, full-stack and mobile products, and the cloud platform underneath them. I am co-founder and CTO of Body.Scratch, where the whole technical stack is mine to run, and ZinZinTech is my own registered studio, where I design, build and ship apps end to end.
Open to part-time engineering roles in the Netherlands. No sponsorship required.
About
AI engineer with six years of international experience across six countries. I work at three layers of the same problem: the models, the product around them, and the infrastructure they run on. Most recently that has meant computer-vision models serving live traffic, a Java and Python backend, and a Google Cloud platform I took over and have run single-handed since. I also run ZinZinTech, my own registered studio, where I build and ship the apps I want to exist.
What I do
Computer vision and NLP models that serve real traffic: pose estimation and gesture classification behind gRPC and TensorFlow Serving, gradient-boosted models scoring live sessions, and an LLM assistant for teachers whose per-student access checks sit inside its own tool layer.
I took over a Google Cloud platform and have run it single-handed since: fourteen Terraform stacks, a GKE cluster, nine Helm charts of my own, and the self-hosted middleware on top of it — RabbitMQ, Keycloak, LiveKit, TensorFlow Serving. Cost is part of the job: spot pools, node pools that scale to zero, committed-use discounts, and a billing kill-switch I wrote.
Production backends in Java with Quarkus and Python with FastAPI, React and TypeScript front-ends, and one Flutter codebase shipped to five platforms through six separate store review processes.
Building datasets when none exist: capture tooling for synchronised multi-camera recording, annotation pipelines, and terabytes of video uploaded to object storage with a queue carrying the recording jobs. Earlier, Spark, Kafka and Elasticsearch behind credit-risk and fraud models we built and evaluated.
Experience
Body.Scratch B.V. · Eindhoven, Netherlands
The technical side is mine end to end: the Google Cloud platform, the production API, the computer-vision stack, and the data pipeline behind it. My co-founder runs business and design.
Commercial client, under NDA
Video copyright-detection system: transformer embeddings indexed for similarity search over a reference library, with temporal alignment to locate the matching segment. Ran on the client's on-premise Windows GPU server, with real users.
Fondazione Bruno Kessler (FBK) · Trento, Italy
Thesis: Trajectory-based Anomaly Detection for Video Surveillance Analysis. Detecting people in video, then clustering their trajectories to flag anomalous movement. The work was merged upstream into the PROTECTOR video-analysis platform.
PipePredict · Darmstadt, Germany
Map-based web application for predictive maintenance of water pipe networks.
Tima Group · Hanoi, Vietnam
Led 5 engineers building OCR and chatbot systems plus credit-scoring and fraud models.
JAVIS · Hanoi, Vietnam
Automated 2D and 3D CAD generation with image processing.
NTU Singapore · MAGIC
I built both halves of a dynamic 3D city-streaming system: the streaming server (PHP, Yii2) and the Unity client — 120 of that repository's 132 commits — a RabbitMQ consumer inside the game engine, loading asset bundles asynchronously and evicting them as the camera moved away. CityGML, level-of-detail scheduling, Berlin and Amsterdam. Two publications, one best paper.
SoICT, Hanoi University of Science and Technology · Vietnam
Combinatorial optimisation for vehicle routing and share-a-ride problems — GRASP and adaptive large neighbourhood search, in a Java optimisation library. Two publications, one as first author, presented by me.
Nakagawa Laboratory ↗, Tokyo University of Agriculture and Technology · Japan
Deep learning for recognising handwritten characters in historical Japanese books, in Prof. Masaki Nakagawa's handwriting-recognition group.
Code
Public repositories. The private ones — the Body.Scratch and ZinZinTech application source, and client work under NDA — are not linked here, but I can walk through any of them.
Reimplementation of Gunasekara et al., Findings of EMNLP 2021: reinforcement learning with question-answering rewards for abstractive summarisation. GPT-2 and PEGASUS on XSum, trained on a Slurm cluster. One model scoring another model's output — and the score means nothing until it agrees with a person.
k-means in C and C++ with hybrid MPI and OpenMP: 14.1x speed-up at 88% efficiency on 16 processes, timed with MPI_Wtime.
Rasa dialogue system over live football data — fixtures, team information and player statistics from an API, with speech input.
Domain-adaptive neural machine translation experiments, with the Slurm batch scripts that ran them.
Transformer-based text recognition, trained on a university GPU cluster.
Significance testing for comparing scene-text-recognition models: is the gap between two systems real, or is it the test set? Built on the Scene Text Telescope reference implementation, credited in the repository.
One of seven detection codebases I trained and fine-tuned in 2022-2023 on a shared university cluster: Slurm jobs that build the model's custom CUDA operators inside the job itself, fine-tune from a released checkpoint, and log to Weights & Biases. The model and the original code are the upstream authors'; the README says exactly what I changed.
A merged pull request in a public repository I do not own: trajectory-clustering anomaly detection integrated into the PROTECTOR video-analysis platform. 621 lines across seven commits, merged by the maintainer in January 2023. My MSc thesis work, landed in someone else's codebase.
Products
Focus, backed by science.
A Pomodoro timer and task planner that helps you focus without guilt. Grow a zen garden as you work, see where your time really goes — and keep every byte of your data on your own device.
Learn programming by dancing.
An AI-powered immersive game for 8-16 youth to learn programming by dancing. Helping educators engage their learners playfully, healthfully and collaboratively.
bodyscratch.academy ↗
Skills
Education
Research
Cloud-Based Dynamic Streaming and Loading of 3D Scene ↗
IEEE Cyberworlds 2018
Efficient Scheduling for Dynamic Streaming of 3D Scene for Mobile Devices ↗
IWAIT-IFMIA 2019 · Best Paper
Traveling Salesman Problem with Multiple Drones ↗
SoICT 2018, ACM
Contact
I am open to part-time engineering roles in the Netherlands — machine learning, software, or platform and DevOps. The fastest way to reach me is email.
phan.business.nl@gmail.com