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
dskit — Reproducible ML project scaffold
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.
credit-risk-classifier — Credit risk scoring system
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.
stock-market-prediction — Time series under real constraints
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.