Welcome to Braedyn Thompson's portfolio!
See the roles I'm targeting →

Now playing
Berkeley Lab SeismicSoCal BearLM
PART-TIME · MENTORING · AI & ML

AperioHub

Research Mentor in AI & Machine Learning for AperioHub Advisory: a one-to-one, project-based research engagement with a motivated high school student, online, built to end in a concrete, defensible result.

Oct 2026 – PresentRemote · Singapore HQ · US–AsiaMentoring · Research methods · AI/ML
8–10 sessions · ~1 h each · weekly · online12345678910REVIEW PROGRESS›NEW CONCEPT›FOCUSED TASKevery session recorded, with AI-assisted notes

Overview01

1:1Mentoring formatone student, one mentor
8–10Sessions~1 hour each, weekly, online
5Research directionsML systems · neural nets · open data · LLMs · AI ethics
1Deliverablea concrete, defensible research output

The aim is for the student to finish with one research result they can explain and defend.

AperioHub Advisory (Singapore, UAE and India) runs project-based experiential learning for students and early-career professionals alongside its consultancy work. As a Research Mentor I guide one high school student through a focused research direction in AI and machine learning, from choosing a question to defending the result.

  • 01

    Mentor a high school student one-to-one through a structured 8–10 session AI/ML research engagement (online, roughly one hour a week, scheduled across US and Asia time zones): each session reviews progress, introduces one new concept or direction and assigns a focused task, so the student ends with a concrete, defensible output.

The direction is chosen with the student in the first sessions; it isn't fixed yet.

ML systems & model evaluationpick a baseline worth beating, hold out data the model never saw, report the gap with its uncertainty
Neural networksbuild a small model from the ground up, then ask what it actually learned and where it fails
Applied AI on open datasetsone question asked of a public dataset, with the cleaning and leakage checks that come with it
Large language modelsevaluate prompts, retrieval and grounding on a fixed question set
AI ethics & bias analysismeasure a bias claim on data, with a method that someone else could re-run

How a session runs02

Every session runs the same loop:

01Reviewwhat was done since last time, and what it showed
02Introducea single new concept or a change of direction
03Assigna focused task to finish before the next session

Sessions are recorded and AI-assisted notes are generated automatically, so the student has a reference for each session and a running record of the whole project, including the dead ends.

What does the student walk away with at the end?

A question they chose, a method they can explain, a result they can defend, and the record of how they got there.

So the write-up is what they show for it?

Yes, and it includes the limits of the result.

What I bring to it03

I teach the loop I use in my own work: a reproducible pipeline, a baseline worth beating, evaluation on data the model has never seen, and an honest write-up of the limits.

  • Baselines and held-out tests: every SeismicSoCal claim sits next to a classical baseline, a confidence interval and a replay of real days.
  • Pre-registration: the BayRidership causal study and the agent A/B test were written down before any result was seen.
  • Knowing when the answer is a ceiling: at CBU the useful result was that the data, not the model, was the limit.
  • Explaining models from the inside: the PatchTST explainer and the adaptive-engine write-ups are the kind of plain-language method sections I ask for.