Skills: Python, R, SQL, NumPy, Pandas, Scikit-learn,
XGBoost, LightGBM, TensorFlow, Keras, PyTorch, HuggingFace,
Transformers, Spark MLlib, MLflow, Kubeflow, Docker, AWS.

Core skills:
• Python (5 yrs) — production ML services, not just notebooks
• XGBoost — 3 shipped models (fraud, churn, demand forecast)
• PyTorch — 1 production NLP model (NER), retrained weekly
• MLflow — model registry across 14 versions; CI gate on eval AUC

- Built a fraud detection model with 0.92 AUC.

- Fraud model (XGBoost): 0.92 AUC vs 0.81 baseline rule
  engine on a 6-month backtest. In production, false-
  positive rate dropped from 1.1% to 0.4% — manual review
  queue down 43% (~$28k/qtr saved).

- Worked on feature engineering.

- Built 42 features across 6 sources via Spark; added
  leakage checks on 12 time-dependent features; moved
  to a Feast feature store shared with 3 models.

- Built a churn model in a Jupyter notebook.

- Shipped churn model as a FastAPI endpoint (p95 < 80ms)
  behind an internal auth gateway; 2M req/day; nightly
  batch scoring via Airflow for email triggers.

- Monitored model performance.

- Added drift monitoring (PSI + KS on 12 features);
  alerts fire when PSI > 0.2 for 2 consecutive days.
- Weekly retraining via Airflow; 14 model versions
  tracked; rollback to previous version in <2 min.

Objective:
Seeking an ML Engineer role where I can leverage my ML
and cloud skills.

ML Engineer · 4 yrs
Python · XGBoost · FastAPI · MLflow · Spark
Shipped 3 production models; fraud model cut manual
review queue 43% (~$28k/qtr saved).

ML Engineer — Nimbus Labs (2021-2024)
- Trained ML models using Python and Scikit-learn.
- Worked on fraud detection and churn prediction.
- Used Pandas for data cleaning.
- Deployed models on AWS.
- Worked with product teams.

Projects
- MNIST digit classifier.
- Titanic survival model.

ML Engineer — Nimbus Labs (2021-2024)
- Owned 3 production models end-to-end (features → training
  → deployment → monitoring → quarterly retrain).
- Shipped a fraud model as a FastAPI endpoint (p95 < 80ms)
  serving 2M req/day; 0.92 AUC vs 0.81 baseline rule engine;
  manual review queue down 43% (~$28k/qtr saved).
- Built a churn model on 40k accounts (3 years); time-based
  holdout; 0.84 AUC vs 0.71 logistic baseline; top-decile
  segment 3.2× churn risk; 2.1 pp churn reduction in Q3.
- Built 42 features across 6 sources via Spark; added
  leakage checks on 12 time-dependent features.
- Added drift monitoring (PSI + KS on 12 features); alerts
  fire when PSI > 0.2 for 2 consecutive days.
- Registered 14 model versions in MLflow; CI gate blocks
  deploys when eval AUC drops >2% vs prod.
- Ran a 2-week A/B (50/50 split, ~120k users) vs. the rule
  engine; +6.2% CTR with no guardrail regression.

Projects
- Recommendation model on 12M interactions (public dataset)
  — +18% CTR vs popularity baseline; public notebook + repo.
  github.com/you/recs-model

Modelling checklist:
☐ Every model has a baseline (rule, logistic, heuristic).
☐ Every metric paired with business impact.
☐ Dataset scale + holdout protocol named.
☐ One feature pipeline or feature store mentioned.
☐ One training cost or inference latency number.
☐ No bullet copied from a job description.

Operations checklist:
☐ One deployment line (API, batch, streaming).
☐ One MLOps line (MLflow, CI gate, registry).
☐ One monitoring/drift line (PSI, KS, alerts).
☐ One A/B or online evaluation.
☐ One retraining cadence or rollback story.
☐ One p95 latency or throughput number.

Positioning checklist:
☐ Opening line = role + stack + one impact number.
☐ One "owned end-to-end" statement.
☐ Stakeholders named (product team, risk, growth).
☐ Decisions moved (what changed because of the model).
☐ One page for 2-5 years of experience.
☐ Filename: Firstname-Lastname-MLE-Resume.pdf