I build production-ready machine learning systems — from raw data to deployed models — designed to be reproducible, testable, and operational in real environments.

Most ML projects fail outside the notebook. I focus on production constraints: data quality, reproducibility, evaluation correctness, and system reliability.

Background in systems administration (ASIR). I design ML systems with infrastructure, failure modes, and performance in mind before model complexity.

• End-to-end ML pipelines (ingestion → validation → feature engineering → training → evaluation → inference)
• Data pipelines with schema validation, quality checks, and leakage prevention
• Reproducible environments with versioned data and locked dependencies (uv, Docker)
• Calibrated probability outputs (Brier score, reliability curves), not raw model scores
• Time-aware evaluation frameworks (walk-forward / stratified CV depending on problem type)
• ML systems decoupled from business decision logic
• Batch inference pipelines designed for scheduled production workloads

Certified: Microsoft Azure Data Fundamentals · Power BI (DAX)
Credly

Production-grade ML template designed to eliminate environment drift and enforce consistent project structure.

Why it matters
Most ML failures are not model failures — they are reproducibility and data consistency failures.

What it enforces

• Strict project structure (data/, features/, models/, pipelines/)
• Dependency locking with uv
• Documentation system with Sphinx
• Pandas / Polars interoperability

Result
Faster setup, consistent engineering standards, zero environment ambiguity.

ML system designed for real decision-making, focused on calibrated probabilities instead of raw predictions.

Key decisions

• Logistic Regression + Random Forest for interpretability vs performance trade-off
• Probability calibration (Platt scaling / isotonic regression)
• Decision threshold decoupled from model (business layer owns decision policy)

Evaluation

• Stratified k-fold cross-validation
• Brier score + AUC as primary metrics
• Strict leakage prevention across time and folds

Result
AUC: 0.81 with calibrated outputs suitable for operational decision systems.

ML applied to a non-stationary, low signal-to-noise environment under realistic constraints.

What most people do wrong
Random splits → leakage → inflated performance

What this project enforces

• Walk-forward validation (deployment simulation)
• Baseline-first evaluation discipline
• Strict no-leakage constraints

Result
Marginal improvement over baseline, consistent with efficient market behavior.

I design ML systems that remain stable under real-world constraints: shifting data distributions, imperfect labels, and production latency.

I don’t optimize notebooks. I design systems that survive production.

I enjoy learning new technologies and adapting quickly to different problem domains. I’m comfortable working across the full ML stack and iterating on systems from prototype to production.