RTX AI & ML research platform

Fields: GPU computing · applied machine learning · MLOps · research tooling

The idea

My HP Z440 is more than a server for web applications: it's a place to develop, execute, instrument, and repeat machine learning experiments on my own hardware. The GPU is an NVIDIA RTX 4060 Ti with 16 GB of VRAM.

Instead of scattering datasets, outputs and environment notes across notebooks, I designed a reusable experimental workflow that connects local computation with versioning, artifact storage, orchestration and monitoring.

Components and responsibilities

Component	What it does
Ubuntu + NVIDIA GPU	Local compute and acceleration
Python + JupyterLab	Prototyping and experimental code
Docker Compose	Repeatable supporting services
DVC	Track datasets and experiment inputs
MinIO	Store generated datasets, checkpoints, and experiment artifacts
Prefect	Organize experiment tasks and dependencies
Custom Python SDK	Consistent experiment logging and instrumentation
Prometheus + Grafana + Loki	Observe system load, application behavior, and logs

The components above describe the research stack and its intended roles. They are not claims that a particular experiment achieved a specific speed, accuracy or uptime.

Example experimental flow

Dataset / DVC revision
       ↓
Experiment config + Python environment
       ↓
Prefect task flow → RTX / CUDA workload
       ↓
Metrics, logs and model artifacts
       ↓
MinIO storage + experiment notes
       ↓
Review performance, change one variable, repeat

What I want an experiment to answer

Which model and dataset revision produced the result?
What exact code and configuration did it run with?
How long did execution and inference take?
How much GPU memory and CPU capacity did the job require?
Where are the artifacts, logs and evaluation results?

Keeping these answers together makes small experiments more valuable and easier to reproduce.

Local inference and the public demo

I am preparing an isolated RTX inference playground so visitors can explore a local model through a simple UI. The model-serving application is separate from the internal MLOps stack and the workstation's administration tools.

What comes next

Publish a complete reproducible run with its dataset revision, task graph, artifact manifest and measured latency/throughput. Until that evidence is published, the site intentionally avoids presenting benchmark numbers as achieved results.

See also: MLOps workflow and architecture.