Research & active projects.

Competing in ROGII's wellbore geology prediction competition — a machine-learning challenge from a Houston-based oil & gas software company. The task is to predict the geological position (True Vertical Thickness) along the lateral section of horizontal wellbores from sparse training data.

Why it matters (and why we're in it)

Subsurface prediction is the same class of problem that shows up in many of our own product areas: making confident judgments from sparse, irregular, high-stakes data. The skills that win a Kaggle competition like this — careful cross-validation, ensemble construction, calibration discipline — are the same skills that make our shipped products reliable.

Approach (high-level)

• Geometric reformulation. The problem reduces to predicting a 2D spatial surface, then projecting back to the original space — a reframing that simplifies what initially looks like a sequence-modeling task.
• Ensemble of disjoint method families. Spatial statistics (universal kriging), neural surface models, and sequence-attention models combined with careful per-row clipping to prevent catastrophic outliers from dragging the average.
• Variance reduction. Seed bagging across stochastic components and adaptive weighting based on local prediction confidence.
• Bias-anchored residual learning. Recent iterations decompose the prediction into a deterministic kriging anchor plus a learned residual — letting the neural net focus on what kriging can't capture.

What we're learning

The biggest insights — careful calibration of cross-validation vs. leaderboard metrics, variance reduction for ensembled neural models, and the discipline of resisting fragile-but-impressive single-method "wins" — apply directly to our shipped products. Reliability under distribution shift is a feature you don't notice until it breaks, and it doesn't break by accident.

The product you're currently on. A real-time food truck discovery app, built mobile-first by a one-person studio.

Some of the most interesting engineering inside GrubTrucks isn't visible to the end user: how we cache third-party APIs without violating their terms, how we avoid storing location data on the backend, how we keep the cryptocurrency QR feature outside the regulatory perimeter for money transmitters. The conservative discipline that shows up in our Kaggle work shows up here too.

Our first shipping product. An AI-driven Mac disk cleaner that uses contextual judgment instead of generic rules — knowing that the same iOS simulator runtime is dead weight for a writer and load-bearing for a developer.

The same posture that drives our Kaggle approach — careful judgment, honest calibration, resistance to over-confident single-method wins — drives MacJanitor's product behavior. Every proposed deletion is shown to you before it runs. Conservative by default. The conservative pass is the trust-building one.

Magpie is structured to ship a portfolio of focused utilities — each one applying careful AI and statistical judgment to a chore that currently requires deep manual expertise. Specifics on what's next will be announced when there's something to actually show.

If you're an indie operator, food truck owner, Mac power user, ML practitioner, or somebody who appreciates one-person studios that ship — keep an eye on this page and on magpiestudios.app.