Data Engineer
See the work →PDF ↓Build pipelines that ingest messy sources reliably and reproducibly.
Welcome to Braedyn Thompson's portfolio!
See the roles I'm targeting →
I build machine-learning systems end to end: the data pipeline, the model, the evaluation, and the product people use.
The roles below are where I want to work next. Each one links to the projects and internships that show the skills it needs.
Build pipelines that ingest messy sources reliably and reproducibly.
Design honest experiments and find out what the data can support.
Turn raw logs and metrics into clear charts and decisions.
Train models that beat a real baseline, then run them in production.
Run experiments carefully and report the limits along with the results.
Build grounded LLM features and measure how well they work.
Make models retrain, promote and monitor themselves, with the gates written down.
Ship full products: backend services, APIs and desktop apps, tested and deployed.
I like problems where I can check whether the answer is right.
That check is the first thing I set up. Before I build, I work out what a good answer would look like and how I would know I had one: a baseline worth beating, a held-out set the model never saw, and a number I can stand behind. Most of what follows is there to make that number trustworthy.
My projects run across forecasting, seismology, retrieval and ed-tech, and they follow the same loop. Build a reproducible pipeline, pick a baseline, test on data the model has not seen (chronological splits, ablations, placebo tests), then ship the result as something a person can open and use.
I care about the pipeline as much as the model, because a forecast is only as good as the data under it. So I spend real time on the unglamorous parts: ingestion, cleaning with a logged reason for every dropped row, dbt models with a test on every build, and orchestration that reruns itself on a free VM while I am asleep.
Sometimes the honest answer is that the data cannot support the question. At CBU I spent weeks showing that an inverse-FEA predictor was held back by its inputs rather than the model, and the team stopped tuning and started fixing the measurements. I would rather report a ceiling than a number I cannot defend.
I build the product end too, since a model nobody can open is not finished. SeismicSoCal sends a real push alert when it detects a quake, BayRidership answers questions through a guarded SQL agent, and JobbyMail runs entirely on your own computer. Shipping forces the parts I would otherwise skip: latency, failure states, and what the thing does when the input is strange.