CHRISTIAN
CALLAHAN
Business Intelligence · Dashboards
I build the dashboards executives use to find the problem: the few numbers leadership acts on, one agreed definition behind each, and the drill from a red flag to the cause. Four years as the BI analyst inside a critical access hospital, where that view moved patient satisfaction 22 points and replaced a $150,000 vendor implementation. SQL and warehouse modeling underneath, Tableau, Streamlit, or Next.js on top.
What I Build
Executive dashboards, and the analysis that makes them worth opening.
Executive First
Built for the person who has to decide: the few numbers that carry a decision, drillable from the red flag to the department, the driver, and the date.
One Definition
One agreed definition, owner, formula, and source per metric, so finance and operations walk into the same meeting with the same number.
Honest Numbers
Real baselines, calibration, and stated limitations. Synthetic and real data kept separate, every metric committed and runnable from a clean clone.
Stack
Portfolio
A few things I've shipped
Rural Hospital Closure Risk
M.S. capstone: 1–2 year closure-risk prediction for U.S. rural hospitals from public CMS cost reports — 89,912 hospital-years (1997–2021) containing just 133 closures, a 0.148% base rate. XGBoost, discrete-time hazard, and logistic models under forward-chaining temporal CV, isotonic calibration, SHAP, and a 10-check leakage audit. Hold-out AUROC 0.867 with 43.7× lift; 3 of 6 real closures rank in the top 1.2%. The central finding is a defect story: an in-sample calibration bug collapsed scores to ~0 for 11,759 of 11,760 hold-out rows — a false negative (AUROC 0.50) initially blamed on COVID relief. 156 tests.
ED Operations Analytics
Site-level forecasting of NHS Scotland A&E 4-hour compliance on Public Health Scotland open data (7,022 Type-1 site-months, 2007–2026). Chronological split, frozen config, holdout scored exactly once; a DuckDB star schema reconciled row-for-row against the Python pipeline. The ensemble beats the persistence baseline 2.72pp vs 2.87pp MAE — and the README leads with the paired-bootstrap CI on that improvement including zero. 111 tests.
A/B Test & Experimentation Analyzer
A decision engine that returns ship / hold / iterate / kill, not a p-value: two-proportion tests, power vs a pre-specified minimum worthwhile effect, Wald CIs, Cohen's h, sample-ratio-mismatch checks, and Holm-corrected exploratory segments. On the canonical 290,584-user experiment it returns a well-powered null — p=0.19 with >99% power to detect the 1pp effect worth shipping. Every README figure regenerates from one script; 16 analytical assumptions documented. 37 tests.
SignalForge
Logistic regression vs. random forest vs. gradient boosting on IBM Telco (7,043 customers), with leak-free cross-validation, bootstrap 95% CIs, paired t-tests, and calibration. The models land within ~0.003 AUC with overlapping CIs, so the writeup treats selection as a calibration/interpretability decision rather than an accuracy contest.
Ticket Intel
Routing and extractive summarization on Banking77 (77 intents) using TF-IDF + Naive Bayes by design: fast, cheap, interpretable, with the router abstracted so a transformer can drop in later.
Pit Wall Intelligence
Ingested 4 seasons of lap-level F1 data (85 races, 90k laps, 33 circuits) through a DuckDB + dbt warehouse. Trained an isotonic tyre-degradation model (1.38s within-circuit MAE; 9.4s leave-one-circuit-out median) and a calibrated LightGBM undercut classifier (AUC 0.66 ± 0.05 on 5-fold GroupKFold, Brier 0.084). Validated a Monte Carlo race simulator against 3 famous 2024 strategy calls (Monaco / Hungary / Italy); average MAE 1.65 finishing positions. Shipped a 6-page Streamlit dashboard, a containerized FastAPI inference service (17ms median latency), and a weekly automated retraining workflow with MLflow tracking.
Churn ROI Simulator
Time-windowed (observation / gap / check) labeling on RetailRocket (2.76M events, 1.41M visitors); LightGBM (Optuna-tuned, isotonic-calibrated) vs. a logistic baseline, plus a budget-targeting ROI simulator. Honest result: the baseline wins the holdout (0.91 vs 0.83); CV 0.88 ± 0.06; calibration cut the test Brier score from 0.065 to 0.009.
Ecommerce Retention & Growth
30-day churn prediction on the WSDM KKBox dataset: calibrated XGBoost (PR-AUC and calibration emphasized under ~9% churn), K-Means LTV segmentation, and a retention-ROI simulator. Ships a synthetic generator so the pipeline runs without the large download.
Healthcare SQL Analytics
Production SQL patterns for EHR analytics on Meditech Paragon, written against the reporting cycles hospital teams actually run: wRVU physician productivity, SDOH quality measures, sepsis bundles, and 340B compliance — the backend of four years as a hospital's in-house BI analyst.
Experience
Four years inside the hospital, now doing the same work on retainer
Founder & Principal
CGC Labs
2026 - Present
Executive dashboards and BI, direct and as a subcontractor to consulting firms
Executive dashboards that show VPs and the C-suite where the problem is: the few numbers leadership acts on, one agreed definition behind each, and the drill from a red flag to the cause. Built on whatever the client already runs (Tableau, Python/Streamlit, Excel) against existing EMR and finance sources.
Key Impact
one view per audience, drillable from the flag to the department, the driver, and the date
driver analysis that names the cause behind a stuck number, priced and assigned to an owner
the BI workstream under a consulting firm's badge: white-label deliverables, NDA/MSA/IC in place, HIPAA and PHI insured
Business Intelligence Analyst
Community Hospital (Critical Access)
2022 - 2026
McCook, NE
Owned the BI function at a critical access hospital: the executive dashboards the C-suite reported to the board, and the analysis underneath them. Replaced a failed $150K vendor solution with custom Tableau and SQL infrastructure. Did the BI-side data transformation on the Veradigm-to-Paragon EMR migration, alongside Altera. Built the reporting that survived CMS audit.
Key Impact
first-ever 75th percentile HCAHPS ranking for the facility
custom Tableau/SQL system delivering $10K/yr in ongoing savings
BI-side data transformation on the Veradigm-to-Paragon cutover, alongside Altera; reporting kept intact
quality, risk, and operational analytics embedded in clinical workflows
Education
Dual MBA & M.S. Data Science
Eastern University
Expected 2027Bachelor of Applied Science
Peru State College
2022Recent Activity
Continuous learning and shipping.
Latest Commits
Loading commits...
Current Focus
Rural Healthcare Analytics
Closure-risk prediction for rural hospitals on CMS cost data, ED-operations forecasting, and the EHR SQL patterns underneath — the domain where my four years of hospital BI live, now with models on top.
Experimentation & Honest Evaluation
A/B testing that returns decisions, not p-values: power vs a pre-specified MDE, SRM checks, and Holm-corrected segments — alongside the calibration and leakage-safe validation running through the churn and forecasting work.
Connect
Open to data scientist and healthcare analytics roles. Email is fastest; the code is on GitHub, the history is on LinkedIn.
christian.g.callahan@gmail.com