Pfizer Round 4 — Day-Of Supplement · Veeva / Orthogonal Data / Semantics

PFIZER ROUND 4 — DAY-OF SUPPLEMENT · Veeva Fluency / Orthogonal Data / Data Semantics · April 15, 2026 · 2:30 PM

New intelligence integrated from all three LinkedIn profiles · Supplement to CheatSheet + Strategic Intel docs · Use all three together

30 min, one interviewer · Elnaz is the decision maker · 80% decided · Confirm fit, don't defend
Interview chain context: Round 2 = John Pastor (IT Dir) · Round 3 = Shetty/peers · Round 4 = Elnaz (her vision)

★ VEEVA — SHE BUILT IT FROM THE INSIDE. THIS CHANGES EVERYTHING.

What you now know: Elnaz was not just a Veeva user. She was Head of Data Science, Commercial Analytics at Veeva Systems for 4 years (2020–2024). She built HCP prioritization models, customer segmentation, and propensity-to-prescribe models ON the Veeva platform using IQVIA/Symphony Health data. Then she came to Pfizer as a buyer and consumer of the same platform. She knows Veeva analytics from both sides of the table.

The trap: If you talk about Veeva at a surface level — "CRM system, stores HCP data" — she will immediately know you're bluffing. She built the analytics layer there. Match her level or stay silent.

The opportunity: She made a deliberate decision to move from Veeva (vendor) to Pfizer (client). She knows what Veeva analytics delivers — and what it doesn't. She joined Pfizer to build something that Veeva's off-the-shelf models couldn't do well enough. That tension is your opening.

Veeva CRM + Analytics — what she built at Veeva, what she uses at Pfizer

Layer	What it is	Elnaz's lens
Veeva CRM	HCP/HCO relationship data — calls, samples, reach/frequency, channel preference	The commercial engagement record of truth. Every AI model she runs at Pfizer starts here.
Veeva Vault (Medical)	Content management, MLR approval workflows, medical documents, CTD submissions	Medical affairs and regulatory content lifecycle — different from commercial analytics but adjacent
Veeva Link	HCP/HCO reference data, affiliation mapping, KOL network intelligence	The identity layer — who is this physician, what institution, what prescribing patterns. Essential for HCP targeting.
Veeva Nitro / OpenData	Prescriber analytics, secondary data (IQVIA, Symphony), integrated with CRM activity data	THIS is what she built at Veeva. Propensity models sat on top of Nitro + IQVIA scripts data.
Veeva Compass	Integrated patient/prescriber longitudinal data — prescriptions, diagnosis, procedure claims	Patient-level claims data feeding care gap logic. She moved from building this at Veeva to consuming it at Pfizer.

How to talk about Veeva WITH Elnaz — at her level

What she knows that matters: Veeva's commercial analytics — propensity scores, reach/frequency optimization, HCP segmentation — are powerful for commercial engagement, but they're built on prescribing patterns, not clinical outcomes. Care gap analytics requires patient-level data (claims, lab, EMR) that the CRM layer doesn't natively carry. She crossed from Veeva to Pfizer to solve that gap.

The pivot she made: At Veeva she optimized commercial engagement (how Pfizer reps reach HCPs). At Pfizer she's optimizing clinical outcomes (whether the right patients are being identified and treated). Her AI testing need is not CRM testing — it's outcome analytics testing. That's a different level of accountability.

Say this if Veeva comes up: "The Veeva layer gives you HCP engagement signals — reach, frequency, channel preference. The care gap logic sits upstream of that, in the patient data. The testing challenge is that the AI tools touching care gap identification are making inferences about clinical unmet need, not just commercial activity. That's where error consequences are different — a CRM recommendation that misses is a lost call. A care gap model that misses is a patient who doesn't get to medication."

Round 2 + 3 context — why you're here

Round	Interviewer	Their world	Why you passed
R2	John Pastor Dir, Business Technology (16 yrs Pfizer IT)	Global Data Mgmt Solution Engineering for BioPharma. Infrastructure, system integration, data pipelines. Not analytics — plumbing.	Demonstrated you could work inside their technical infrastructure and governance model. IT credibility + 50 min + 12 min overage = strong buy signal.
R3	Rachael Rathbun + peers (Shetty: MedConnect Data Lake, 148 sources, $35M budget)	Commercial analytics delivery, data governance, enterprise data acquisition. They OWN the data pipelines feeding Elnaz's models.	Showed cultural fit. They work downstream of Elnaz — they need a tester who won't break their pipelines or add noise. They endorsed you TO Elnaz.
R4	Elnaz Alipour, PhD Decision maker	Care gaps, HCP prioritization, patient segmentation. She defines what "correct" looks like for AI outputs. Her team endorsed you.	Not decided yet. She's confirming judgment fit. ~80% made. Confirm it.

Key framing: Shetty owns the MedConnect Data Lake — 148 data sources, 15+ global markets. That infrastructure is what feeds Elnaz's AI models. When you talk about data quality testing upstream of AI outputs, you're speaking to the world Shetty described and Elnaz depends on.

Orthogonal Data — what it means in Elnaz's world

Definition (in her context): Orthogonal data sources are independent, non-overlapping data inputs that each capture a different dimension of the same patient or HCP — when combined, they produce a complete picture that no single source can provide alone. Each source is "orthogonal" because it answers a different question, with minimal redundancy between sources.

Elnaz's orthogonal data stack for care gap modeling:

Claims data (Rx/medical) — what was prescribed, filled, and billed. Prescribing behavior, adherence proxies.
Lab/diagnostic data — what was clinically measured. Biomarker levels, disease severity, uncontrolled condition signals.
EMR/EHR data — what the physician documented. Diagnosis codes, treatment history, notes.
Pharmacy/specialty data — what was actually dispensed. Hub enrollment, PAP utilization, specialty fill rates.
Patient-reported outcomes — patient self-report. Symptom burden, quality of life, adherence self-report.

The testing problem with orthogonal data: When you join orthogonal data sources, coverage is not uniform. EMR data is denser for commercially-insured patients than Medicaid. Specialty pharmacy data misses patients using retail fill. A care gap model that ingests all five sources will systematically undercount care gaps in populations with sparser data coverage — even if the AI logic is perfect. That's a data completeness bias, not an AI failure — and it's invisible unless you test across population subgroups.

Say this: "The challenge with orthogonal data is that completeness isn't uniform across patient populations. If the model is trained and validated on commercially-insured patients where EMR and claims data are dense, it will under-identify care gaps in populations where one or two data sources are sparse. That's a testing vector I'd add to the baseline suite — care gap identification rate parity across data coverage tiers."

Data Semantics — the definition consistency problem

Definition: Data semantics is the assurance that a business term means exactly the same thing across every system, model, query, and report that uses it. In care gap analytics: "care gap," "uncontrolled patient," "HCP priority," and "eligible patient" must have identical definitions in the source SQL, the AI model prompt, the model's training data, and the output delivered to the commercial team.

Semantic drift — the silent failure mode: Semantic drift occurs when the same term is defined slightly differently in different systems. Example: "uncontrolled hypertension" defined as SBP > 140 in the claims-derived care gap rule, but the Cortex AI was fine-tuned on literature where "uncontrolled" means SBP > 130. The model identifies a larger patient population than the intended definition. The commercial team acts on an inflated care gap list. No one failed — the definitions just diverged.

Why this hits Elnaz's world specifically: She spent 4 years at Veeva where commercial analytics definitions varied by client. At Pfizer she's now responsible for the definitions that feed downstream decisions. When she evaluates AI tools, she needs to know: does this system's understanding of "care gap" match our internal operational definition exactly? If it drifted during fine-tuning or in-context prompt construction — the outputs are wrong in a way that's invisible to users.

Testing approach — semantic consistency testing:

Extract the canonical definition of each key business term from the approved data dictionary or SQL rule set.
Test whether the AI produces the same patient population as the SQL-defined rule for that term.
Any patient in AI output not in the SQL result = semantic overreach. Any patient in SQL result not in AI output = semantic underreach. Both are semantic failures.

Say this: "Semantic consistency testing is a separate test vector from hallucination testing. The AI can return a factually grounded response that still violates the business definition — it used 'care gap' the way a clinical paper defined it, not the way Pfizer's operational ruleset defines it. I'd validate each key term against the data dictionary definition with a SQL ground truth comparison."

★ 30-MIN STRATEGY — EVERY MINUTE COUNTS

Min 0–2 (Opening): Use the collaborative opener from the Strategic Intel doc. Don't re-pitch the resume — she has it. Reference Rachael's conversation: "I have a clear picture of what you're building." Establish you're here to confirm fit, not audition.

Min 2–20 (Her questions): This is hers. Answer directly. No preamble. Every answer ends with a bridge to her domain: care gaps, HCP targeting, or segmentation model integrity. If she asks technical — go technical. She's a physicist. She won't punish depth.

Min 20–25 (Your questions): Ask 2 max. Lead with the Vision question from the Strategic Intel doc: "What does the analytics function look like when the AI testing is working correctly — what changes about how your team operates?" Then the Governance question if time. These show you've thought at her level, not the BA level.

Min 25–30 (Close): "What I'd bring immediately is the ability to protect the integrity of the outputs your team depends on — with documented evidence that holds up in a regulated environment. I understand those outputs ultimately feed HCP and patient decisions, and that testing them is a data integrity function. That's the role I want."

Say-this-not-that — Veeva, Semantics, Orthogonal

Veeva is a CRM system that pharma companies use to track HCP interactions

→

Veeva's CRM captures commercial engagement signals — reach, frequency, channel preference — but the care gap logic sits in patient-level claims data upstream of that. Different accountability level.

I'd test the AI output against the expected answer

→

I'd extract the canonical SQL definition of each term from your data dictionary and validate whether the AI produces the same patient population. Delta = semantic failure, not hallucination.

I'd test across different types of patients

→

I'd stratify care gap identification rate by data coverage tier — patients with all five sources vs. those with only two. Differential error rates across coverage tiers reveal model dependence on data completeness, not AI quality.

I understand pharma from my pharmacy background

→

The commercial analytics domain — HCP prioritization and care gap segmentation — is the layer I'm building depth in specifically to serve your team's tools. My pharma background gave me the clinical safety frame; the commercial analytics frame is what I'd be adding in this role.

I'm a fast learner and can pick up whatever tools you use

→

The Snowflake Cortex stack — Analyst, Search, Agents, AI Observability — is the exact environment I've been preparing for. Cortex Agents are the complex case: failures are emergent across the reasoning chain. I'm ready for that.

Watch-outs — 30 minutes is not enough to recover from these

Do NOT ask her to explain Veeva. She built it. You'll lose credibility in 10 seconds. Speak at her level or bridge past it.
Do NOT present yourself as learning pharma commercial analytics. She needs someone who can operate inside her world immediately. Acknowledge it as a depth addition, not a starting point.
Do NOT over-explain the QA background. She already bought it through rounds 2 and 3. She's confirming judgment fit, not re-evaluating credentials.
Do NOT ask more than 2 questions. 30 minutes. She has questions. Protect her time and she'll respect yours.
Do NOT sell on overqualification. If she raises it: "The BA title is the entry point into pharma commercial AI testing at the right org. I'm optimizing for domain, not title."

If she asks about Shetty's MedConnect data lake

What you know: Jayaprakash Shetty's team built MedConnect — 148 data sources, 15+ global markets, direct ingestion for Medical Affairs. eLAAD integrates open claims data. These are the data pipelines that feed Elnaz's AI models. Shetty is the delivery lead — he owns the infrastructure you'd be testing against.

Say: "From what Rachael described, the data infrastructure here is sophisticated — multi-source ingestion across global markets. That's exactly the environment where upstream data quality testing has the highest leverage. If the AI output is wrong, I need to be able to distinguish: is this a model failure, a prompt failure, or a data pipeline issue? That diagnostic precision is the value I'd bring."