Novant Health  ·  AI Technical Product Architect II  ·  Round 1  ·  2026-05-21  ·  2:00 PM EST  ·  30 MIN  ·  Teams Video  ·  jd_id=37

Interviewer Intel — Read Both Profiles Cold

Person Type What They Actually Built / What They Will Test Bishop Bridge — Say This Specifically
Mayur Kotadia
Director AI Platform
11 yr Novant. ETL Lead → Architect → Supervisor → Manager → Director Data Integration (Oct 2022) → Director AI Platform (Nov 2024). Azure cert. Cloud Governance Committee member. ETL: Epic Clarity ETL, Epic Caboodle ETL, Azure Data Factory, Informatica, SSIS, Attunity, Synapse. Built Novant's cloud operating model. Explicitly built QA/testing standards for ETL teams.		 HM
PRIMARY
DECISION		 He will ask:
• Can you operate inside a governed data platform, not just build models on top of it?
• Do you understand Epic Clarity/Caboodle data structures at the field level?
• Can you architect at the platform layer, not just the model layer?
• Will you respect and extend existing governance, not bypass it?
• Do you understand Azure Data Factory and Synapse as the pipeline backbone?
He is a data platform engineer who just became an AI director. His lens: platform stability and governance first, AI second. 		"I reverse-engineered 147 Epic Clarity and Caboodle reports at Evernorth — field-level, source-to-target mapping, null-handling variations, semantic inconsistencies across workflows. I built the governed KPI library and data dictionary that became the AI foundation. That's the same problem your platform team owns." "At Invistics I built a vendor-agnostic canonical data model that ran a single ML model across 8+ different EHR vendors and 3+ ADC vendors across 300+ facilities. Enterprise data integration at scale, not a single-system implementation." "I'm a LEAN Six Sigma Black Belt. QA standards for data pipelines — I've built these. I understand why you formalized testing/QA processes for your ETL team."
William Espinoza
Data Science Manager
4 yr 8 mo Novant. GT BS Industrial Engineering + MS Health Systems. Prior: Children's Healthcare of Atlanta, GT research. Key production project: Primary Aldosteronism Risk Model — CLARITY source, Databricks/PySpark on billions of rows, custom MLflow training classes with hyperopt, Logistic Regression (outperformed XGBoost/LGBM on small cohort <2k), AUC 0.79/AUPRC 0.31/ECE 0.06, bias assessment sex/race/age, 30k predictions into Epic daily. Model Surveillance POC: SPCC techniques. Also: Surgical Demand Forecasting (SARIMAX, Prophet, LSTM). Prior: OPTUM commercial claims 100GB in SAS, CCI calculation, logistic regression.		 TV
TECH
VALIDATOR		 He will ask:
• Do you understand statistical validation rigor — not just AUC but calibration, bias assessment, assumption testing?
• Can you talk to MLflow specifically — experiment tracking, model registry, custom training patterns?
• Have you actually pushed predictions into Epic in production — not just built models?
• What is your approach to model drift monitoring? (He used SPCC — that's Six Sigma.)
• How do you handle small clinical cohorts where gradient boosting fails?
He's a rigorous applied statistician who codes. He will probe whether Bishop understands the gap between model training and clinical production deployment. 		"I'm using MLflow right now at NewsRx for experiment tracking, model registry, and version control across 11 iterations. At Optinosis I'm building the MLflow model registry infrastructure for the SEER-Medicare ML pipeline. MLflow is not abstract to me." "I'm a LEAN Six Sigma Black Belt. SPCC for model drift monitoring is standard Six Sigma tooling. The Model Surveillance POC approach you've used is the right architecture — control charts on prediction distributions, thresholds set with clinical input. I've been doing SPC in regulated environments since nuclear pharmacy." "At Invistics my production pipeline fed clinical decisions at 300+ facilities for 7 years. The validation architecture: pharmacist-labeled gold standard, population-stratified accuracy, clinician feedback loop. The same pattern you use on the Primary Aldosteronism model — domain expert co-ownership of ground truth."

Key Novant Facts to Drop (Confirm You've Done Homework)

  • 35,000 team members, $10.2B revenue, 800+ locations
  • Institute for Innovation & AI founded 2019
  • Director AI Platform role is new (Nov 2024) — Mayur's promotion
  • DAX Copilot: 900 clinicians, 550K+ encounters (Microsoft/Nuance + Epic)
  • ML stack: Azure + Databricks + MLflow + PySpark
  • Data lake: Azure Synapse + Starburst/Trino
  • Clinical AI: Aidoc (imaging), Viz.ai (stroke), Lung Cancer AI
  • CitiusTech partnership (Sept 2024) — decision intelligence
  • HIMSS Level 9/10 Digital Health — top 1% globally
  • 2 open roles — ask which track (RevCycle vs Clinical) early

Opening Pitch — Mixed Panel (Lead Outcome → Technical Credibility)

When asked "Tell us about yourself" — 45 seconds, tight:
My work has been in production for nine years. Five systems. Zero failures. At Invistics I took an NIH-funded algorithm, built a vendor-agnostic canonical data model that ran across 8 EHR vendors and 3 dispensing cabinet vendors, and commercialized it to 300+ facilities at 96% accuracy — peer-reviewed in AJHP. That system runs today. At Evernorth I reverse-engineered 147 Epic Clarity and Caboodle reports to resolve the semantic data issues blocking AI — built the governed KPI library and data dictionary that became the AI foundation. Right now I'm delivering production GenAI at NewsRx — Qwen3-32B on self-hosted vLLM, MLflow experiment tracking, 21 CFR Part 11 audit trail, 83% preference over human-written content. I'm also hands-on weekly as architect and data scientist at Optinosis, building an early lung cancer detection system for Medicare Advantage using SEER-Medicare ML. I'm a LEAN Six Sigma Black Belt. I hold SaMD certification from Mayo Clinic. I have a Master of Divinity in Ethics, which is not a common credential for this seat. I understand you're scaling the AI Platform — I'd like to know what's constraining that scale right now.

Top Walkthrough Q&A — 60-Second Answers

Q [Mayur]: Walk me through how you've dealt with Epic data inconsistencies that blocked AI implementation.
At Evernorth I had 147 Epic and Tableau reports — each one built by different analysts over different years, using different field mappings, different KPI definitions, different null-handling assumptions. Before any model could run, that semantic layer had to be governed. I reverse-engineered every report source-to-target, built a standardized data dictionary, and created a governed KPI library. That became the data foundation the AI runs on. The Epic Clarity and Caboodle layers have specific patterns: Clarity is transactional, Caboodle is analytical but with its own versioning inconsistencies per implementation. The field I've seen break the most AI models is the admit/discharge/transfer date — handled differently across facilities, not consistently normalized in the ETL. You build governance around what actually breaks, not a theoretical data model.
Q [William]: Walk me through your statistical validation approach for a clinical ML model.
At Invistics: pharmacist-labeled ground truth for the gold standard — domain expert labeling, not crowd-sourced. Population-stratified accuracy across facility types and patient demographics before production approval. Calibration check: AUC tells you discrimination; ECE tells you whether the model's confidence is trustworthy — you need both. For clinical cohorts under 2,000 records, standard logistic regression typically outperforms gradient boosting because calibration matters more than marginal discrimination gains on small samples. I confirm statistical assumptions before trusting the model: Cook's Distance for outliers, VIF for multicollinearity if you're running logistic. Drift monitoring in production using SPCC — I'm a LEAN Six Sigma Black Belt, Statistical Process Control is standard tooling in my quality practice. That's how you catch drift before it becomes a patient event.
Q [Both]: What does responsible AI mean when the stakes are clinical outcomes?
Three things that cannot be retrofitted. First: the domain expert is a co-designer, not a sign-off. At Invistics, clinical pharmacists labeled edge cases from day one — that's why the model was trusted at 300+ facilities. Second: every prediction has traceable lineage. At NewsRx I run ALCOA+ audit trails and 21 CFR Part 11 electronic signatures on every scored output — not because a regulator required it, but because that's what production integrity requires. Third: the model is advisory, the licensed professional decides. That human-in-the-loop architecture is the governance. I have a Master of Divinity in Ethics. The ethical architecture is not an afterthought — it's the foundation. My most stark example: I designed the original Gracie AI for Street Grace — national CSEC deterrence, operational since 2017. 30+ states, 140+ law enforcement agencies, 78+ cities. Georgia beta alone: 54,000+ messages exchanged, 13,000+ predators reported — the single largest verifiable study on sex trafficking ever conducted. Each disruption is one transaction that never happened, one potential victim who was never exploited. When the stakes are a child's safety, you cannot skip governance. That discipline carries into every clinical system I build.

Critical Watch-outs — 30 MIN HARD STOP

⚠ TIME — TEK Carrier feedback: answers ran long. 60-sec max. After each: "Did that answer your question?"
⚠ POC TRAP — Mayur flagged this from screening. Every answer lands on: deployed → scaled → measurable outcome. If you catch yourself saying "POC," pivot immediately to what it became.
⚠ AZURE GAP — Don't volunteer it. If asked: "My production work is AWS and GCP. Azure ML architectural patterns are equivalent — ADF maps to Glue, Synapse maps to Redshift/BigQuery. I'd ramp on Novant-specific configurations fast; the governance model is what I'm solving, not the console."
⚠ OVERQUALIFICATION — Defuse early if the energy shifts. "I'm on keyboards at two active builds right now — NewsRx and Optinosis. Not advising. Building."
⚠ TRACK UNKNOWN — Ask within first 2 minutes: "The JD mentions two open roles — Revenue Cycle and Clinical. Which track is this conversation focused on?" Your answer framing changes based on the response.
⚠ DON'T DISCUSS — Rate, remote terms, availability, benefits. TEK negotiates rate. Darbee said explicitly: don't go there.

What Each Interviewer Is REALLY Testing

Mayur: "My entire 11-year career is this data platform. Can you operate within it without breaking governance or creating shadow infrastructure? Do you understand that the AI runs on top of the data layer — and that data layer is mine?" He wants: partnership, not a cowboy who bypasses his pipelines.
William: "I push 30k predictions into Epic every day. I wrote custom MLflow training classes. I use SPCC for surveillance. Is Bishop a production data scientist or a strategy consultant who talks about models?" He wants: technical peer, not a manager who delegates code.

SVPG Vocab to Use Naturally

Unique Credentials — Drop These If Natural

FRONT PAGE  |  Novant Health AI Technical Product Architect R1  ·  2026-05-21  ·  4Shadow AI