TL;DR:
I work in data and run a home lab for home automation, learning, and hosting local data and ML pipelines. I’m building a fully local system that lets users ask natural-language questions of structured data and machine learning models and receive clear explanations back. My current RTX 3060 limits model size and token throughput, and an R9700 would allow me to run larger models, longer contexts, and more complex, interpretable workflows.
I’m a data engineer by trade with a strong data science background. Over the last several years, I’ve worked extensively with SQL-heavy analytics pipelines, machine learning workflows, and production data systems for small companies. Much of my work focuses on building systems that are understandable, maintainable, and practical to use.
Outside of work, I run a fairly involved home lab where I self-host, among far too many other projects and services, end-to-end data and machine learning pipelines. These include ingesting, processing, and modeling datasets related to things like local real estate, weather and time-series data, and especially sports analytics. I use my home server to experiment with local inference, vector search, and reproducible data workflows while keeping everything private and under my own control.
Recently, I decided to delve into self-hosting an LLM to explore the privacy and convenience of having this capability available locally. I was able to afford a used RTX 3060 12GB from a local marketplace and started hosting Ollama with smaller models. This quickly led me to think about how LLMs could be integrated more deeply into my existing self-hosted workflows to help solve problems I was already encountering.
The recurring problem: interpretability
As I build more model-driven systems, one issue comes up repeatedly (especially when trying to share my ‘exciting cool projects’ with my wife and other less technically-enthused audiences): interpretability.
I do not just want an answer from a black box, I want to be able to understand and explain in conversational terms why the result was produced, which data was used, or how the logic maps back to SQL queries, aggregations, or transformations
This is especially important when working with large, real-world datasets like sports statistics, housing data, or long-running time series. Outputs that cannot be traced back to the underlying data are difficult to trust and hard to improve.
The project I’m working on: Ask the Data
This project is my attempt to address that gap.
I’m building a fully local system that allows users to ask natural-language questions about structured datasets and machine learning models, receive a clear explanation of the result, and ask follow-up questions that build on prior context. For more technical users, the system is designed to surface the logic behind the workflow rather than hiding it.
The goal is to make working with data more approachable without obscuring how results are derived.
Sports analytics is the initial focus because it provides a challenging but well-scoped problem space. It involves large historical datasets, context-heavy questions, and a strong need for explanation rather than opaque outputs. My current sports models focus on MLB and NBA/NCAA basketball data.
Some example questions I’m targeting include:
- Why does an NBA lineup perform worse later in games?
- What changed statistically when this player was traded and which team was more positively affected?
- How does this team perform against left-handed pitching over time, and what factors contribute to that?
- If the model gives the Spurs an 83% chance to win tonight, what factors are driving that result?
The focus is on structured analysis and explanation, presented in a way that anyone, regardless of skill or technical background, can understand and use.
Why local compute and why the R9700 matters
Everything in this project is designed to run locally. There are no cloud APIs or external services involved. This makes it easier to experiment, iterate, and understand system behavior end to end.
At the moment, my main constraint is hardware.
I’m currently using an RTX 3060, which has been sufficient for early prototyping with very small models and datasets. As both increase in size, however, I quickly run into time and memory constraints. Token throughput drops as model size or context length grows, and multi-step reasoning, longer explanations, and iterative follow-up questions become slow and can even crash the inference process due to VRAM limits.
An R9700 would allow me to:
- Run larger local models with longer context windows
- Perform faster embedding and retrieval over larger datasets
- Support multi-step queries and conversational follow-ups at usable speeds
This would make it possible to move beyond a constrained prototype and continue developing the project in a more meaningful way.
Where this is going
While sports analytics is the initial focus, the underlying system is intentionally flexible. The same approach applies to data within local real estate, weather, and other structured datasets and models where understanding how an answer was produced is just as important as the answer itself.
I plan to document progress, design decisions, and lessons learned as the project evolves. Interpretability and explainability remain significant challenges in modern machine learning, and I’m interested in building tools that help bridge that gap, both within my home lab and for others working on similar problems.