Practice Interview Questions

Edge → specific project → bridge to this role anchored to a JD line → honest gap with closing plan. See 02 §story.

Specific. "Internal-only feedback loops are too forgiving for me. I do my best engineering when the deadline and audience are both real. FDE puts me there week to week." Avoid generic "I love working with people."

Domain interest matters. Strong: "Enterprise spend is a problem domain where the data has been illegible for decades — large companies don't know what they're contractually obligated to. Making that legible is the kind of problem I want to spend years on." Avoid surface enthusiasm.

STAR. Lead with the decision you owned (not the framework). Include the alternative you considered. Measurable outcome.

Have a story. Frame: vague customer ask, your scoping moves (clarifying questions, working agreement, what you deferred), the artifact you shipped, what changed downstream.

Underrated; most candidates don't have this story. Having it differentiates. What was being built, why it became clear it wouldn't ship usefully, what you documented, how you reallocated. Lead with what you learned about the customer's problem in the kill.

Direct, evidence-based, with an alternative. "I disagreed because [data]. I proposed [alternative]. They went with [outcome]." Capitulation reads junior; obstinance reads worse. The senior shape is "no, and here's the better thing."

You over-promised, misunderstood, or hit something you didn't see coming. How you reset. The shape is: name the gap quickly, propose a new commitment that's defensible, communicate it cleanly, deliver on the reset.

Pre-deployment quality sample. Pull 50–100 representative docs covering the claimed format mix. OCR-quality check by eye. Run extraction on a subset and compare to manual reads. Quantify pass rate, HITL rate, failure rate. Calibrate customer expectations before signing SLA, not after.

Five steps. Internal alert before customer notices. Diagnose (drift, regression, customer data change, miscalibration). Contain (push more to HITL, tighten threshold). Notify customer proactively. Root-cause (file platform bug, update config, adjust HITL guidance). Don't declare victory until rolling window is back above SLA.

Auditability (customer disputes a value — point at source page), debugging (which region did the model look at), and selective reprocess (re-run only docs with suspect provenance instead of everything).

Six core entities. Suppliers (with parent-supplier hierarchy), documents (physical), agreements (logical, multiple documents per agreement), document-agreement links (M:N), obligations (extracted commitments, versioned), spend (linked to agreements when possible). Separate document and agreement; canonicalize suppliers; store raw + resolved names.

You'll reprocess constantly — new model versions, new doc classes, bug fixes. Without idempotency, reprocesses duplicate or require manual cleanup. With stable IDs (content hash + customer), upserts, and versioned outputs, reprocess is routine. Document IDs derive from content hash; extractions are keyed by (document_id, model_version).

Scope out loud first. Ask: Friday hard? 12 formats equally important or do 3 cover 80%? Cheapest useful artifact? Then sequence: ingest top-3 formats by Wednesday, extraction with tight HITL on representative sample, initial output to staging warehouse by Thursday, defer long-tail formats and SLA-grade quality to next week with explicit caveats. Status doc Friday morning. Senior signal: naming what to defer and why.

Understand what they actually need (often different from what they asked). Check if existing capability gets 80% of the way. If yes, ship the 80% version with explicit caveats. If no, decide if it's platform-team territory — file with customer evidence. Either way, agree in writing on what they're getting, what they're not, and the gap-closure timeline.

Diagnose with evidence first — quality sample, specific examples. Bring to customer with data, not opinion. Propose alternatives: re-scope to the data that works, push pre-processing onto their side (re-scan, normalize), or relax SLA carve-outs. Don't pretend it's fine. The honest conversation in week 3 saves the deployment from failing in week 12.

Discovery. Read 20+ contracts, talk to data users (analysts), data consumers (often different), executive sponsor (different again). Map existing stack. Find the manual workflow you're replacing. Don't code. Don't promise outcomes. By end of week 2, one-page working agreement signed off by senior stakeholder.

Push back first. "What decision is real-time enabling that hourly wouldn't?" 80% of "real-time" asks are actually "fresh enough to act on today." If genuinely real-time, evaluate webhook push vs streaming — and make the operational cost (on-call, monitoring, complexity) visible to the customer before committing.

Map stack, file identity tickets (Okta SSO + SCIM) day 2 — IT is the long pole. Design read paths first: SAP supplier master, Coupa contract refs. Agree Snowflake landing schema with data team. Scope Looker semantic layer with analysts. No code in week one; ticketing and discovery only.

Political (system of record has weight), risky (bad write affects close, payables, audit), sequencing (most value lands in warehouse and BI without touching ERP). Build that value first, earn the right to write back.

Shadow against production 1–2 weeks. Promote to staging schema; customer's data team validates. File CAB ticket (if required) with risk, validation, rollback, comms. Deploy in maintenance window. Monitor SLA metrics. Revert if any class regresses. Document customer-facing notes.

Six sections on one page: symptom, severity, triage (first 10 min), likely causes (in order of frequency), resolution paths, escalation. Written for the customer's data analyst, not for you — vocabulary they know, links they have access to, no internal jargon. Test by handing it to a real analyst who wasn't involved in the deployment.

Three stages over months 3–6. You primary → joint (customer first-responder following your runbooks) → customer-primary (you available for platform escalations only). Transition criterion: their team can resolve top-5 most common issues without paging you. Test with intentional drills.

MSA establishes the overarching commercial and legal relationship (liability, indemnification, IP, payment). SOWs sit underneath an MSA and define specific engagements (scope, deliverables, pricing, timeline). One MSA can have many SOWs. Separate documents, linked into one logical agreement.

Spend with no active agreement covering it — 15–30% of enterprise spend typically. Routing it back to contracts captures negotiated pricing, ensures policy compliance, reduces operational risk. Often the headline ROI number platforms in this space pitch.

Procurement leadership (CPO / VP Procurement) usually leads, co-buyers in finance (Controller / VP Finance) and legal (General Counsel / VP Legal). Each cares about different facets: procurement = supplier visibility / consolidation; finance = spend accuracy / audit prep; legal = risk surfacing.

Auto-renew clauses with missed notice windows lock customers into year-long extensions at often-higher prices. Catching upcoming renewals with notice deadlines lets the procurement team renegotiate, cancel, or consolidate in time. ROI is direct and measurable, which is why it's the lead pitch.

Don't ship into political stalemate. Surface to your AE / leadership. Either escalate internally on the customer side (get senior alignment) or pause the deployment until they align. Shipping conflicting outcomes wastes your time and damages the relationship. Naming this as a "we'll pause until your team agrees" move builds trust over time.

Stakeholder is the customer not internal product. Done means customer can act and stakeholders agree, not spec met. Scope is what you find, not what you're given. Failure mode is relationship + technical, not just bug. Cadence is weekly with the customer, not sprint internal. Performance review is customer expansion + deployment durability, not pure velocity. The biggest day-to-day difference: knowing when to stop engineering and start talking.

Don't list infrastructure. Lead with the question: week 1 in customer conversations + corpus reading; week 2 a working agreement signed with the senior stakeholder; weeks 3–4 the first end-to-end slice (one source, one document class, output landing in staging warehouse). Senior signal: starting with the question, not the architecture; one signed agreement before any code.