Gibson AI

Independent ML & Analytics Consultant

Structured Machine Learning Systems from Operational Data

Greg Gibson

I design reproducible pipelines and disciplined evaluation frameworks that turn operational metrics into predictive insight.

Discuss a Project View Case Study Field Notes on AI
  • Pipeline architecture
  • Reproducible experimentation
  • Threshold-aware evaluation
  • Leakage prevention
Professional headshot of Greg Gibson

Featured project

A flagship ML build that emphasizes structure, evaluation, and reproducibility.

CardioSentinel

A structured ML system for threshold-aware heart attack risk modeling.

Case Study
Domain Healthcare risk prediction
Constraint Class imbalance (~36% positive)
Focus Precision ↔ recall tradeoffs

Problem

In risk modeling, overall accuracy can mask critical failure modes. A model may appear strong while missing the cases that matter most.

Approach

  • Config-driven pipeline factory (repeatable experiments)
  • Safe vs. learned feature separation to prevent leakage
  • MLflow tracking for structured model comparison
  • Precision-floor thresholding to align evaluation with use case

Outcome

Above a precision floor of 0.40, recall declined sharply for several models. A balanced logistic regression maintained more stable behavior across thresholds, suggesting a largely monotonic signal. In this setting, a well-regularized linear model matched or outperformed boosted trees while remaining easier to interpret and calibrate.

View on GitHub Discuss a similar build

This threshold-first evaluation approach applies across healthcare, fraud, churn, and safety modeling.

Consulting

I help teams turn operational metrics into reliable analytics and reproducible ML pipelines — so results hold up outside the notebook.

Operational analytics → predictive insight

Clarify the metric definitions, build a clean dataset, and design an evaluation plan that reflects real tradeoffs (false positives vs. false negatives, cost, capacity, and risk).

  • KPI definitions & data sanity checks
  • Target/label design and leakage review
  • Decision thresholds tied to use-case

Reproducible pipelines & automation

Convert ad-hoc analysis into maintainable workflows: consistent preprocessing, feature governance, and repeatable runs you (or your team) can operate confidently.

  • Pipeline refactors (notebook → modules)
  • Leakage-aware “safe vs learned” separation
  • Containerized, repeatable execution

Experimentation & model evaluation

Structured comparisons across models and feature sets, with clear artifacts and tracking—so you can explain what changed, why it improved, and what to do next.

  • MLflow tracking (params, metrics, artifacts)
  • PR/ROC analysis + calibration checks
  • Readable summaries for stakeholders

Engagement style: a short diagnostic (data + goal + current workflow), followed by a clear plan. Implementation support as needed. You’ll leave with a system your team can run and maintain—built for clarity, not mystery.

Insights

Practical notes from real modeling work — focused on structure, evaluation discipline, and decision alignment.

Thresholds: when “lower” improves outcomes

Why model performance shifts across decision thresholds — and how to align precision/recall tradeoffs with operational constraints rather than default metrics.

Pipelines vs. notebooks

Where exploration belongs — and how structured pipelines prevent leakage, inconsistency, and silent evaluation errors.

Designing leakage-aware features

Separating safe transformations from learned features to preserve evaluation integrity and maintain trustworthy model comparisons.

Each insight is grounded in applied experimentation and reproducible system design — not theoretical best practices.

Contact

If you’re building, evaluating, or restructuring a predictive system, I’m happy to discuss your data and goals.

Start a Conversation LinkedIn GitHub

What to include: a brief description of your dataset, current workflow, and the decision you’re trying to support. I typically begin with a short diagnostic call and a clear, written plan before any deeper engagement.

Direct email works best. Thoughtful, structured work starts with a clear understanding of the problem.