Chandler Drake Talley — Data Science Portfolio

Senior data scientist who ships production ML and turns results into business impact.
This repo is the evidence: pipelines that avoid leakage, stratified evaluation, SHAP interpretability, and stakeholder-ready recommendations across credit risk, churn, clinical triage, time series, and quant finance. Everything lives here—no external portfolio site.

---

Quick links

This repo	GitHub — code, notebooks, and setup
View notebooks in browser	nbviewer — no install required
LinkedIn (https://www.linkedin.com/in/drake-talley/)	[Add your LinkedIn profile URL]

---

Why this portfolio stands out

Every notebook follows the workflow a senior data scientist uses in production. These were built end-to-end by hand starting in 2020—before LLMs and GenAI—and many were developed alongside a senior data scientist with 30+ years of experience, with his review and approval. They have been refined since with pipelines, SHAP, and current best practices. Recruiters and hiring managers can run the code, see the same metrics, and judge rigor—not just slides.

Practice	Implementation
Reproducibility	set_seed(42), consistent random_state across all models and splits
No data leakage	sklearn.Pipeline wraps scaler → selector → estimator; fit only on train folds
Stratified splits	stratify=y on every train_test_split and StratifiedKFold CV
Multi-metric eval	cross_validate with accuracy, precision, recall, F1, ROC-AUC simultaneously
Interpretability	SHAP TreeExplainer/KernelExplainer on every supervised project
Business framing	Each project ends with stakeholder-ready recommendations and cost analysis
Threshold tuning	Precision-recall tradeoff analysis with cost-sensitive optimization
Time series	Daily trend, rolling means, day-of-week/hour seasonality, temporal heatmaps; TimeSeriesSplit and lag features where appropriate

---

Projects at a glance

Project	One-line impact
Corporate Bankruptcy	Predict bankruptcy risk from 96 financials → credit exposure mitigation.
Telecom Churn	Predict churn → cost-sensitive retention (e.g. $500 acquisition vs $75 offer).
Heart Disease	Predict from biometrics → clinical triage with sensitivity-first metrics and calibration.
NJ Transit + Amtrak	Predict delays from 98K trips → schedule padding, crew allocation, passenger alerts.
NYC Bus	Cluster 6.7M bus records → segment-specific scheduling and anomaly detection.
Jane Street	Predict profitable trades → position sizing and transaction-cost-aware signals.
RAG + LLM	FAISS + Sentence Transformers + local LLM → no-api-key demo.

Full project descriptions and tech stacks are in Projects below.

---

For recruiters & hiring managers

5-minute tour:

1. Open Modern_Classification_Workflow_Bankruptcy.ipynb — end-to-end pipeline, CV, SHAP, and “why” in one place.
2. Run streamlit run app.py — interactive EDA, model comparison, SHAP, and live predictions.
3. Skim docs/BEST_PRACTICES.md — explains reproducibility, pipelines, and interpretability choices.

Skills demonstrated (mapping to typical job descriptions):

• ML engineering: sklearn.Pipeline, ColumnTransformer, stratified CV, hyperparameter tuning (GridSearchCV / RandomizedSearchCV).
• Model deployment: FastAPI serving endpoint with /predict, /predict/batch, /explain (SHAP), Dockerfile, and docker-compose.
• Interpretability & compliance: SHAP (TreeExplainer/KernelExplainer), feature importance, threshold tuning for cost-sensitive decisions.
• Evaluation: Multi-metric reporting (precision, recall, F1, ROC-AUC), calibration for probability outputs, time-series-aware splits.
• Data engineering: DuckDB-backed SQL analytics alongside pandas; production-pattern for warehouse-style EDA.
• Testing: pytest suite for utilities, data loaders, API endpoints, and training pipeline.
• Production hygiene: Reproducible seeds, locked envs (pyproject.toml / requirements.txt), smoke tests, Docker.
• Business communication: Each project ends with stakeholder recommendations, quantified ROI, and cost/benefit framing.

---

Model Serving API (FastAPI)

The Telecom Churn model is deployed as a production-style REST API with cost-sensitive predictions and SHAP explanations.

# Train the model (saves artifacts to artifacts/)
python -m api.train

# Start the API
uvicorn api.serve:app --reload

# Or run everything in Docker
docker-compose up

Endpoints:

Method	Path	Description
GET	/health	Health check
GET	/model/info	Model metadata, metrics, feature names, cost assumptions
POST	/predict	Single customer churn prediction with risk tier and action
POST	/predict/batch	Batch predictions with summary statistics
POST	/explain	SHAP-based explanation of top churn drivers

Example:

curl -X POST http://localhost:8000/predict \
  -H "Content-Type: application/json" \
  -d '{"gender":"Male","SeniorCitizen":0,"Partner":"Yes","Dependents":"No",
       "tenure":12,"PhoneService":"Yes","MultipleLines":"No",
       "InternetService":"Fiber optic","OnlineSecurity":"No","OnlineBackup":"Yes",
       "DeviceProtection":"No","TechSupport":"No","StreamingTV":"No",
       "StreamingMovies":"No","Contract":"Month-to-month","PaperlessBilling":"Yes",
       "PaymentMethod":"Electronic check","MonthlyCharges":70.35}'

---

Interactive Dashboard

streamlit run app.py

Features per project:

• EDA tab — target distributions, feature correlations, interactive drill-downs
• Model comparison — train multiple models with sklearn Pipelines and stratified 5-fold CV
• SHAP tab — TreeExplainer feature importance plots
• Live prediction — adjust feature sliders and see real-time predictions with confidence scores

---

Setup

1. Virtual environment (recommended)
   Use the project venv so notebooks run with a consistent interpreter:

   python -m venv venv
   .\venv\Scripts\Activate.ps1
   pip install -r requirements-core.txt   # notebooks: pandas, sklearn, xgboost, jupyter, etc.
   pip install -r requirements.txt        # full stack: + tensorflow, RAG (streamlit, torch, transformers)

   In Cursor/VS Code: use the project venv as the notebook kernel so numpy and portfolio_utils are available: Kernel → Select Kernel → Python Environments → ./venv/Scripts/python.exe (or "Python (Data Science Portfolio)" if it appears). This repo's .vscode/settings.json sets that as the default. If you see ModuleNotFoundError: No module named 'numpy', the kernel is using a different Python — run .\venv\Scripts\python.exe scripts/install_deps.py then switch the kernel to the venv. Core deps are in requirements-core.txt; requirements.txt adds TensorFlow and the RAG app deps (install when no other process is using the venv to avoid file locks).

   Alternatively, without a venv: pip install -r requirements.txt. Or uv: uv sync (optional: --extra rag --extra shap).

2. Windows: console warning / timeouts
   If you see RuntimeWarning: Proactor event loop does not implement add_reader or notebooks hang/timeout when running with nbconvert --execute, use the helper script so the correct asyncio policy is set before Jupyter starts:

   python scripts/run_nbconvert.py --execute --inplace .\path\to\notebook.ipynb --ExecutePreprocessor.timeout=1200

   For interactive Jupyter, the warning is harmless; the kernel uses a fallback. If a cell runs for 20+ minutes with no output, stop the kernel and run with the script above or increase the timeout.

3. Download datasets (Kaggle API)

   python setup_data.py --no-jane-street

   Requires a Kaggle API token in ~/.kaggle/kaggle.json. Omit --no-jane-street to include the large competition dataset.

4. Run notebooks — jupyter notebook or open in VS Code/Cursor. Select kernel "Python (Data Science Portfolio)". Data loads via portfolio_utils.data_loader with Colab fallback.

5. Launch dashboard — streamlit run app.py

---

Projects

Corporate Bankruptcy Prediction

Business problem: Predict which companies are at risk of bankruptcy using 96 financial indicators (Taiwan Economic Journal, 1999–2009). Enables creditors to mitigate exposure before default.

Senior analysis includes: Pipeline (StandardScaler → SelectKBest → XGBoost), stratified 5-fold CV, precision-recall threshold tuning for cost-asymmetric decisions, SHAP feature importance, and recommendations for the credit risk team.

Tech: XGBoost, Random Forest, Gradient Boosting, Logistic Regression, SMOTE, SelectKBest, PCA, SHAP

View notebook

---

Telecom Customer Churn (IBM Watson Analytics)

Business problem: Predict which customers will churn so the retention team can intervene proactively. Retaining a customer is 5–25× cheaper than acquiring a new one.

Senior analysis includes: Pipeline with class-weighted models, cost-sensitive threshold optimization ($500 acquisition cost vs. $75 retention offer), SHAP-driven intervention recommendations, and A/B testing strategy.

Tech: XGBoost, Random Forest, Gradient Boosting, Logistic Regression, SVM, SMOTE, GridSearchCV, SHAP

View notebook

---

Heart Disease Prediction (Clinical Decision Support)

Business problem: Predict heart disease from patient biometrics for clinical triage. Missing a case (false negative) can be fatal — sensitivity is the priority metric.

Senior analysis includes: Pipeline-based modeling, calibration curve analysis (critical for clinical probability estimates), SHAP for clinical feature importance, and risk-stratified care pathway recommendations.

Tech: Random Forest, Gradient Boosting, SVM, Logistic Regression, GridSearchCV, SHAP, calibration analysis

View notebook

---

NJ Transit + Amtrak Rail Delay Prediction

Business problem: Predict train delays using supervised, unsupervised, and deep learning on 98K NEC rail trips. Enables proactive passenger notification and resource allocation.

Senior analysis includes: Time series analysis (daily delay trend, 7-day rolling mean, day-of-week and hour-of-day seasonality, day×hour heatmap), pipeline with SelectKBest, multi-metric stratified CV, SHAP for operational insight, and recommendations for schedule padding, crew allocation, and real-time passenger alerts.

Tech: Time series (pandas, seaborn), Decision Tree, Random Forest, Gradient Boosting, KNN, SVM, KMeans, DBSCAN, t-SNE, TensorFlow, SHAP

View notebook

---

NYC Bus Clustering (Unsupervised)

Business problem: Segment 6.7M MTA bus location records into operational clusters for route optimization, service planning, and anomaly detection.

Senior analysis includes: Systematic k-selection with silhouette analysis (not just elbow method), per-cluster silhouette plots, PCA-space visualization with variance explained, and operational recommendations for segment-specific scheduling.

Tech: KMeans, DBSCAN, PCA, t-SNE, silhouette analysis

View notebook

---

Jane Street Market Prediction (Quantitative Finance)

Business problem: Predict profitable trades from anonymized financial features. Precision on trade signals directly impacts P&L.

Senior analysis includes: Pipeline (Imputer → Scaler → XGBoost), stratified CV, SHAP for risk management transparency, and recommendations for time-series validation, position sizing, and transaction cost modeling.

Tech: XGBoost, RandomizedSearchCV, SHAP

View notebook

---

RAG + LLM Demo (Streamlit)

Retrieval-augmented generation with FAISS vector search, Sentence Transformers, and local GPT-2. No API keys required.

streamlit run Streamlit_Langchain_RAG_LLM.py

Tech: Streamlit, FAISS, Sentence Transformers, GPT-2, PyTorch

---

Modern Classification Workflow (Reference)

A clean reference notebook demonstrating the full senior workflow end-to-end: set_seed → Pipeline → stratified CV → multi-metric evaluation → SHAP.

View notebook · Best Practices Guide

---

Testing

pip install pytest httpx duckdb
pytest tests/ -v

Test file	Covers
test_ml_utils.py	Seed reproducibility, pipeline construction, SHAP fallback
test_data_loader.py	Data directory config, CSV discovery, Kaggle slug validation
test_train.py	Preprocessing (encode, drop, target), cost-sensitive threshold optimization
test_api.py	FastAPI endpoints: health, model info, predict, batch, explain
test_db_utils.py	DuckDB loader, SQL queries, context manager, pre-built analytics queries

---

Repo structure

Data-Science-Portfolio/
├── api/                    # FastAPI model serving
│   ├── train.py            #   Train pipeline, serialize artifacts
│   ├── serve.py            #   REST API (predict, explain, batch)
│   └── schemas.py          #   Pydantic request/response models
├── portfolio_utils/        # Shared Python package
│   ├── data_loader.py      #   Kaggle-backed dataset loaders
│   ├── ml_utils.py         #   Seeds, pipelines, SHAP helpers
│   └── db_utils.py         #   DuckDB SQL analytics layer
├── tests/                  # pytest suite
├── docs/                   # Best practices guide
├── scripts/                # Notebook automation (smoke tests, patching)
├── advanced_visualization/ # Dash geospatial dashboard
├── artifacts/              # Trained model + metadata (after api.train)
├── data/                   # Datasets (after setup_data.py)
├── Dockerfile
├── docker-compose.yml
├── pyproject.toml
├── requirements.txt
├── requirements-core.txt
├── requirements-api.txt
├── app.py                  # Streamlit portfolio dashboard
├── setup_data.py           # Download all datasets
└── *.ipynb                 # Project notebooks