You inherit a $3M strategic account with a 12-week deployment window, six customer teams, and no agreed technical sponsor. How would you score discovery and map stakeholders in the first five days?

I would make deployment readiness and decision authority visible before committing engineering capacity. - I would score business value, workflow clarity, data access, integration access, security feasibility, customer staffing, and decision process from 0 to 2, with evidence linked to every nonzero score. - My stakeholder map would name the economic buyer, technical sponsor, workflow owner, data owner, security approver, procurement lead, and likely blocker, plus the decision each one controls. - A zero in data access, security feasibility, or decision authority blocks the build regardless of the total, and I would review that result with the sponsor on day five.

A retailer says it wants an AI assistant to make store operations better within 90 days, but cannot define better. How do you turn that objective into a measurable workflow?

I would narrow the objective to one repeated decision with a baseline, user, input, output, and measurable consequence. - I would shadow store managers handling stockout exceptions and choose a workflow such as producing a replenishment recommendation from inventory, sales, and delivery data. - The target might be reducing median exception handling from 18 to 8 minutes while keeping accepted recommendations above 85% and no unauthorized order submission. - I would document what remains human-owned, which systems supply evidence, and which metric triggers expansion after four weeks of production use.

A bank wants to automate a 14-person loan-review process in ten weeks, but has no reliable process metrics. How would you establish the current-state workflow and human baseline?

I would measure the real work before claiming an automation target. - I would sample at least 100 cases across normal, peak, and exception paths, recording touch time, queue time, rework, handoffs, approval rate, and outcome quality. - I would map the systems and manual artifacts used at each step, including spreadsheet checks and policy lookups that are absent from the official procedure. - The agreed baseline would separate avoidable effort from required judgment, so the deployment is judged against comparable cases rather than an invented average.

A manufacturer expects $5M annual value from a predictive-maintenance deployment, while its platform team estimates 1,200 engineering hours to integrate it. How would you build the ROI map?

I would model gross workflow value and subtract the full customer and vendor cost needed to realize it. - Value would come from avoided downtime, fewer emergency parts, and technician hours saved, each tied to an owner, baseline, adoption rate, and confidence range. - Cost would include the 1,200 customer engineering hours, FDE time, cloud and model spend, security review, training, and ongoing support rather than treating integration as free. - I would publish conservative, expected, and upside cases, then require the expected case to clear the customer's payback threshold before expanding scope.

You are the primary technical owner for a $4M account that must move from discovery to production in 90 days across Snowflake, Okta, and two internal APIs. What does your 30/60/90-day plan contain?

I would organize the plan around evidence gates and usable workflow increments, not three calendar status labels. - By day 30, I want the workflow baseline, architecture, data and identity access, success and kill criteria, dependency owners, and a thin end-to-end path in a nonproduction environment. - By day 60, representative data, failure paths, security evidence, evaluation results, load results, runbook, and trained customer operators must be complete. - Days 61 to 90 cover controlled production rollout, reconciliation, adoption tracking, rollback readiness, and a value review that decides expansion or stop.

A healthcare customer wants a production AI workflow in eight weeks, but identity access takes ten days and representative data has not cleared privacy review. How would you run an integration-readiness review?

I would produce a readiness decision with hard blockers, dated owners, and the last safe start date for each dependency. - The review covers source contracts, sample and production access, data quality, identity, network path, security and privacy approvals, target APIs, customer staffing, and test environments. - I would accept contract-accurate mocks for early adapter work, but they cannot close live data, identity, or network gates. - If privacy approval and identity ownership are not credible by the calculated dates, I would replan the workflow or pause the deployment rather than compress validation.

A 45-day fraud-detection PoC for a $2.5M account needs a custom data adapter and customer labels that may arrive late. What kill criteria would you agree before coding?

I would define stop rules for missing evidence, unsupported engineering, and failure of the core business claim. - Stop if representative labels and source access are unavailable by day 12, because later accuracy results would not leave time for customer review. - Stop if the adapter requires more than 80 approved engineering hours, bypasses supported authentication, or introduces a product fork with no owner. - Stop if the agreed evaluation misses the minimum recall at the allowed false-positive rate after one bounded remediation cycle, and preserve the evidence for a qualified restart.

A 70-day deployment depends on the customer's data, IAM, network, legal, and operations teams plus your product team. How would you define dependencies and RACI without creating meeting overhead?

I would keep one critical-path dependency register with a single accountable owner and decision date for every gate. - Each item names the artifact or outcome, responsible doer, accountable approver, consulted experts, informed stakeholders, predecessor, and last safe completion date. - I would avoid shared accountability: the customer IAM lead owns the SCIM mapping decision, while I own proving our integration against it. - Twice-weekly reviews cover only changed, blocked, or near-trigger items, and overdue dependencies automatically change the forecast or invoke a scope decision.

A logistics customer requests a routing workflow in ten weeks; 70% fits your standard template, 20% needs a supported extension, and 10% appears customer-specific. How do you choose what to build?

I would use the standard template as the default and demand stronger evidence as work moves toward custom code. - Configuration stays in the template when it changes thresholds, mappings, or policy values without altering product behavior. - A supported extension is justified when the interface is stable, isolated, testable, and likely reusable across customers, with an explicit maintenance owner. - I would reject or time-box the final 10% unless it is essential to measured value and its lifetime cost is lower than changing the customer workflow.

At week six of a 12-week deployment, the customer COO adds two regions and a Workday integration while keeping the original launch date. How do you control scope?

I would make the request a decision with visible cost rather than silently absorb it. - I would trace each addition to the signed workflow and value metric, then estimate architecture, security, testing, training, and customer-hour impact. - I would offer options such as one-region launch on time, a bounded Workday stub with a later production gate, or a revised date for the full scope. - The accountable sponsor chooses in writing, and the baseline plan, forecast, kill criteria, and decision log update in the same review.

A fintech stores 12 TB in Snowflake and wants a customer-support AI workflow live in 60 days with source-to-answer freshness under ten minutes. How would you design the pipeline?

I would keep governed data in Snowflake and build an incremental, observable path into the AI workflow. - Snowflake Streams and Tasks or Dynamic Tables produce approved support records with stable IDs, event time, schema version, and ACL metadata through a dedicated least-privilege export role. - A connector batches changed records into the retrieval index, checkpoints progress, quarantines invalid rows, and preserves source version and ACLs so query-time retrieval can enforce access and cite the exact snapshot. - I would measure p95 freshness below ten minutes, reconciliation totals, access-control tests, deletion propagation, retrieval quality, and recovery after a one-hour processing pause.

A media customer has a Databricks lakehouse with 800 million documents and expects a six-week AI search pilot. What integration would you propose?

I would scope a representative pilot corpus rather than imply that all 800 million documents must be indexed in six weeks. - With the customer, I would select document families, languages, access patterns, and real queries that can prove the search decision, while keeping Delta tables as the governed source. - Unity Catalog controls approved columns and service-principal access, and Delta Change Data Feed exposes inserts, updates, and deletes for the pilot cohort after its initial snapshot. - Spark jobs preserve document and version IDs and quarantine rejects; the pilot reports freshness, delete handling, Delta lineage, retrieval quality by cohort, cost per thousand searches, and the estimate for full expansion.

A marketplace will send 60,000 order events per second through Kafka to your platform, and three customer teams consume the results. What event contract do you establish?

I would define a versioned business contract with clear semantic ownership before tuning Kafka. - The envelope includes event ID, order ID, occurred-at time, producer, schema version, correlation ID, and trace context; the payload carries only stable order facts. - Records are keyed by order ID for per-order ordering, schemas enforce backward compatibility, and optional additions pass producer and consumer contract tests. - The customer order team owns meaning, every consumer registers field usage, and deprecation requires a measured migration window rather than a surprise rename.

A mining customer has 15,000 edge sensors sending MQTT over unreliable links and wants anomaly detection within two minutes. How would you design the telemetry path?

I would design for intermittent connectivity, duplicate delivery, and device identity at the edge, while making the two-minute promise conditional on connectivity. - Devices use per-device certificates, scoped topics, QoS 1 for important telemetry, sequence numbers, and a local disk buffer sized for the agreed outage window. - Regional gateways validate schemas, batch and compress messages, preserve event time, and forward to the central pipeline without sharing one fleet credential. - I would agree with operations that p95 sensor-to-decision freshness stays below two minutes for connected devices; offline readings are marked late and cannot trigger stale automated action. - Tests cover a six-hour link loss, duplicate rejection, certificate rotation, backlog catch-up time, and the customer-visible gap report after recovery.

A retailer updates 40 million product rows nightly but wants price changes reflected within five minutes before a 75-day launch. Would you use CDC or batch?

I would use a hybrid: batch for the initial and periodic full baseline, then CDC for price changes that need five-minute freshness. - The bulk load establishes a reconciled snapshot without forcing the source database to stream years of history. - CDC captures inserts, updates, and deletes with source position and schema version, while idempotent consumers apply changes in order for each product. - I would keep a scheduled full reconciliation because CDC can preserve source changes but cannot prove that every downstream transformation remained correct.

A regional insurer offers a REST API, nightly encrypted files, and an internal queue for the same claims data, with a 12-week deployment deadline. How do you choose the integration interface?

I would choose by interaction semantics, freshness, failure recovery, and customer operating cost rather than interface prestige. - REST fits targeted reads or commands needing an immediate response, but I would verify quotas, pagination, idempotency, and availability. - Encrypted files fit high-volume daily snapshots when hours of latency are acceptable and checksums, manifests, and replay are available. - The queue is best for incremental claims events if retention, ordering, identity, dead-letter handling, and customer support ownership are production-ready.

You must migrate three years of support tickets and then ingest live updates without downtime in a 50-day cutover. How do you combine backfill and incremental processing?

I would start the incremental cursor before the bulk copy and join both paths through the same idempotent sink. - I would record the source high-water mark, backfill immutable ranges with stable ticket and version IDs, and retain changes occurring after that mark. - Incremental events can be buffered while the backfill runs, then applied in source order with newer versions winning over stale bulk records. - Cutover requires count and checksum reconciliation, lag at zero, sampled field comparison, and a documented point at which the legacy reader can be disabled.

A lending workflow depends on 25 customer fields, but null rates and code meanings vary by region and the launch is nine weeks away. What data-quality contract do you create?

I would turn data quality into an owned interface with measurable acceptance and quarantine behavior. - For each field, the contract names type, allowed values, semantic owner, freshness, completeness threshold, uniqueness rule, and handling of late or corrected data. - Critical fields such as applicant ID and decision outcome get hard gates, while optional fields can degrade through documented defaults without hiding the loss. - Dashboards show quality by region and source, and threshold breaches route to named customer owners before they contaminate evaluation or production decisions.

A 30,000-user enterprise requires SAML SSO, SCIM provisioning, and contractor separation before an 80-day launch. How would you design identity integration?

I would separate authentication, lifecycle provisioning, and authorization while agreeing one immutable workforce identifier. - SAML assertions use signed responses, strict audience and recipient checks, short validity, and certificate-rotation testing against the customer's identity provider. - SCIM provisions users and groups with pagination, idempotent updates, explicit deactivation, and a target such as access removal within 15 minutes. - Employee and contractor groups map through separate policies, and exit tests cover joiner, mover, leaver, duplicate identity, expired certificate, and failed assertion paths.

A bank has 18 internal roles and your product has six permission bundles; go-live is in seven weeks. How do you map RBAC without creating one-off roles?

I would map customer job functions to the smallest product permissions that satisfy real tasks, not mirror every legacy role name. - A matrix lists each workflow action, sensitive object, customer role, product permission, approver, and segregation-of-duties conflict. - Where several customer roles need the same technical access, I would reuse one product bundle and keep business distinctions in the identity provider groups. - Unmappable privileges become explicit product gaps or denied scope, and tests use representative users to prove both allowed and forbidden actions.

Forward Deployed Engineer interview questions

100 real questions with model answers and explanations for Middle candidates.

See a Forward Deployed Engineer resume example →

Practice with flashcards

Spaced repetition · Hunter Pass

Questions

stakeholder-managementsponsorcommunication

I would make deployment readiness and decision authority visible before committing engineering capacity.

I would score business value, workflow clarity, data access, integration access, security feasibility, customer staffing, and decision process from 0 to 2, with evidence linked to every nonzero score.
My stakeholder map would name the economic buyer, technical sponsor, workflow owner, data owner, security approver, procurement lead, and likely blocker, plus the decision each one controls.
A zero in data access, security feasibility, or decision authority blocks the build regardless of the total, and I would review that result with the sponsor on day five.

Why interviewers ask this: The interviewer checks whether you qualify a strategic deployment through evidence and influence paths rather than enthusiasm alone.

I would narrow the objective to one repeated decision with a baseline, user, input, output, and measurable consequence.

I would shadow store managers handling stockout exceptions and choose a workflow such as producing a replenishment recommendation from inventory, sales, and delivery data.
The target might be reducing median exception handling from 18 to 8 minutes while keeping accepted recommendations above 85% and no unauthorized order submission.
I would document what remains human-owned, which systems supply evidence, and which metric triggers expansion after four weeks of production use.

Why interviewers ask this: A strong answer converts vague executive language into a bounded operational workflow with measurable value and safety.

monitoringconcurrency

I would measure the real work before claiming an automation target.

I would sample at least 100 cases across normal, peak, and exception paths, recording touch time, queue time, rework, handoffs, approval rate, and outcome quality.
I would map the systems and manual artifacts used at each step, including spreadsheet checks and policy lookups that are absent from the official procedure.
The agreed baseline would separate avoidable effort from required judgment, so the deployment is judged against comparable cases rather than an invented average.

Why interviewers ask this: The interviewer evaluates whether you ground value claims in observed customer operations instead of a process diagram supplied by management.

estimationdeployment

I would model gross workflow value and subtract the full customer and vendor cost needed to realize it.

Value would come from avoided downtime, fewer emergency parts, and technician hours saved, each tied to an owner, baseline, adoption rate, and confidence range.
Cost would include the 1,200 customer engineering hours, FDE time, cloud and model spend, security review, training, and ongoing support rather than treating integration as free.
I would publish conservative, expected, and upside cases, then require the expected case to clear the customer's payback threshold before expanding scope.

Why interviewers ask this: A strong answer protects the customer from inflated ROI by pricing engineering effort and adoption into the value case.

snowflakeapi

I would organize the plan around evidence gates and usable workflow increments, not three calendar status labels.

By day 30, I want the workflow baseline, architecture, data and identity access, success and kill criteria, dependency owners, and a thin end-to-end path in a nonproduction environment.
By day 60, representative data, failure paths, security evidence, evaluation results, load results, runbook, and trained customer operators must be complete.
Days 61 to 90 cover controlled production rollout, reconciliation, adoption tracking, rollback readiness, and a value review that decides expansion or stop.

Why interviewers ask this: The interviewer checks whether your deployment plan sequences technical, customer, and value evidence toward a real production outcome.

health-checks

I would produce a readiness decision with hard blockers, dated owners, and the last safe start date for each dependency.

The review covers source contracts, sample and production access, data quality, identity, network path, security and privacy approvals, target APIs, customer staffing, and test environments.
I would accept contract-accurate mocks for early adapter work, but they cannot close live data, identity, or network gates.
If privacy approval and identity ownership are not credible by the calculated dates, I would replan the workflow or pause the deployment rather than compress validation.

Why interviewers ask this: A strong answer distinguishes preparatory work from proof that customer systems are actually ready for deployment.

I would define stop rules for missing evidence, unsupported engineering, and failure of the core business claim.

Stop if representative labels and source access are unavailable by day 12, because later accuracy results would not leave time for customer review.
Stop if the adapter requires more than 80 approved engineering hours, bypasses supported authentication, or introduces a product fork with no owner.
Stop if the agreed evaluation misses the minimum recall at the allowed false-positive rate after one bounded remediation cycle, and preserve the evidence for a qualified restart.

Why interviewers ask this: The interviewer tests whether you can end a low-leverage PoC objectively despite account pressure.

deploymentdependenciesraci

I would keep one critical-path dependency register with a single accountable owner and decision date for every gate.

Each item names the artifact or outcome, responsible doer, accountable approver, consulted experts, informed stakeholders, predecessor, and last safe completion date.
I would avoid shared accountability: the customer IAM lead owns the SCIM mapping decision, while I own proving our integration against it.
Twice-weekly reviews cover only changed, blocked, or near-trigger items, and overdue dependencies automatically change the forecast or invoke a scope decision.

Why interviewers ask this: A strong answer makes cross-company ownership operational without replacing delivery with governance meetings.

I would use the standard template as the default and demand stronger evidence as work moves toward custom code.

Configuration stays in the template when it changes thresholds, mappings, or policy values without altering product behavior.
A supported extension is justified when the interface is stable, isolated, testable, and likely reusable across customers, with an explicit maintenance owner.
I would reject or time-box the final 10% unless it is essential to measured value and its lifetime cost is lower than changing the customer workflow.

Why interviewers ask this: The interviewer evaluates whether you balance customer value with product integrity and long-term ownership.

deployment

I would make the request a decision with visible cost rather than silently absorb it.

I would trace each addition to the signed workflow and value metric, then estimate architecture, security, testing, training, and customer-hour impact.
I would offer options such as one-region launch on time, a bounded Workday stub with a later production gate, or a revised date for the full scope.
The accountable sponsor chooses in writing, and the baseline plan, forecast, kill criteria, and decision log update in the same review.

Why interviewers ask this: A strong answer shows the grit to reset executive expectations while preserving a credible path to value.

snowflakedesignci-cd

I would keep governed data in Snowflake and build an incremental, observable path into the AI workflow.

Snowflake Streams and Tasks or Dynamic Tables produce approved support records with stable IDs, event time, schema version, and ACL metadata through a dedicated least-privilege export role.
A connector batches changed records into the retrieval index, checkpoints progress, quarantines invalid rows, and preserves source version and ACLs so query-time retrieval can enforce access and cite the exact snapshot.
I would measure p95 freshness below ten minutes, reconciliation totals, access-control tests, deletion propagation, retrieval quality, and recovery after a one-hour processing pause.

Why interviewers ask this: The interviewer checks whether you can connect Snowflake to an AI product without losing governance, freshness, or recoverability.

I would scope a representative pilot corpus rather than imply that all 800 million documents must be indexed in six weeks.

With the customer, I would select document families, languages, access patterns, and real queries that can prove the search decision, while keeping Delta tables as the governed source.
Unity Catalog controls approved columns and service-principal access, and Delta Change Data Feed exposes inserts, updates, and deletes for the pilot cohort after its initial snapshot.
Spark jobs preserve document and version IDs and quarantine rejects; the pilot reports freshness, delete handling, Delta lineage, retrieval quality by cohort, cost per thousand searches, and the estimate for full expansion.

Why interviewers ask this: A strong answer combines Databricks mechanics with the customer evidence needed to judge an AI search deployment.

kafka

I would define a versioned business contract with clear semantic ownership before tuning Kafka.

The envelope includes event ID, order ID, occurred-at time, producer, schema version, correlation ID, and trace context; the payload carries only stable order facts.
Records are keyed by order ID for per-order ordering, schemas enforce backward compatibility, and optional additions pass producer and consumer contract tests.
The customer order team owns meaning, every consumer registers field usage, and deprecation requires a measured migration window rather than a surprise rename.

Why interviewers ask this: The interviewer evaluates whether you treat Kafka integration as a durable cross-team contract rather than a transport choice.

design

I would design for intermittent connectivity, duplicate delivery, and device identity at the edge, while making the two-minute promise conditional on connectivity.

Devices use per-device certificates, scoped topics, QoS 1 for important telemetry, sequence numbers, and a local disk buffer sized for the agreed outage window.
Regional gateways validate schemas, batch and compress messages, preserve event time, and forward to the central pipeline without sharing one fleet credential.
I would agree with operations that p95 sensor-to-decision freshness stays below two minutes for connected devices; offline readings are marked late and cannot trigger stale automated action.
Tests cover a six-hour link loss, duplicate rejection, certificate rotation, backlog catch-up time, and the customer-visible gap report after recovery.

Why interviewers ask this: A strong answer addresses the operational realities of customer edge systems rather than assuming a reliable cloud connection.

batch

I would use a hybrid: batch for the initial and periodic full baseline, then CDC for price changes that need five-minute freshness.

The bulk load establishes a reconciled snapshot without forcing the source database to stream years of history.
CDC captures inserts, updates, and deletes with source position and schema version, while idempotent consumers apply changes in order for each product.
I would keep a scheduled full reconciliation because CDC can preserve source changes but cannot prove that every downstream transformation remained correct.

Why interviewers ask this: The interviewer checks whether you choose ingestion modes from freshness and recovery needs instead of declaring one universally superior.

resttypesdata-structures

I would choose by interaction semantics, freshness, failure recovery, and customer operating cost rather than interface prestige.

REST fits targeted reads or commands needing an immediate response, but I would verify quotas, pagination, idempotency, and availability.
Encrypted files fit high-volume daily snapshots when hours of latency are acceptable and checksums, manifests, and replay are available.
The queue is best for incremental claims events if retention, ordering, identity, dead-letter handling, and customer support ownership are production-ready.

Why interviewers ask this: A strong answer ties API, file, and queue choices to the customer's actual workflow and operational constraints.

backfillconcurrency

I would start the incremental cursor before the bulk copy and join both paths through the same idempotent sink.

I would record the source high-water mark, backfill immutable ranges with stable ticket and version IDs, and retain changes occurring after that mark.
Incremental events can be buffered while the backfill runs, then applied in source order with newer versions winning over stale bulk records.
Cutover requires count and checksum reconciliation, lag at zero, sampled field comparison, and a documented point at which the legacy reader can be disabled.

Why interviewers ask this: The interviewer evaluates whether you can avoid data gaps and stale overwrites during a live customer migration.

fundamentals

I would turn data quality into an owned interface with measurable acceptance and quarantine behavior.

For each field, the contract names type, allowed values, semantic owner, freshness, completeness threshold, uniqueness rule, and handling of late or corrected data.
Critical fields such as applicant ID and decision outcome get hard gates, while optional fields can degrade through documented defaults without hiding the loss.
Dashboards show quality by region and source, and threshold breaches route to named customer owners before they contaminate evaluation or production decisions.

Why interviewers ask this: A strong answer treats customer data quality as a shared production contract rather than cleanup performed by the FDE alone.

design

I would separate authentication, lifecycle provisioning, and authorization while agreeing one immutable workforce identifier.

SAML assertions use signed responses, strict audience and recipient checks, short validity, and certificate-rotation testing against the customer's identity provider.
SCIM provisions users and groups with pagination, idempotent updates, explicit deactivation, and a target such as access removal within 15 minutes.
Employee and contractor groups map through separate policies, and exit tests cover joiner, mover, leaver, duplicate identity, expired certificate, and failed assertion paths.

Why interviewers ask this: The interviewer checks whether enterprise identity design covers lifecycle and failure behavior, not only successful login.

rbac

I would map customer job functions to the smallest product permissions that satisfy real tasks, not mirror every legacy role name.

A matrix lists each workflow action, sensitive object, customer role, product permission, approver, and segregation-of-duties conflict.
Where several customer roles need the same technical access, I would reuse one product bundle and keep business distinctions in the identity provider groups.
Unmappable privileges become explicit product gaps or denied scope, and tests use representative users to prove both allowed and forbidden actions.

Why interviewers ask this: A strong answer balances least privilege, customer semantics, and maintainable product authorization.

Locked questions

21
A SaaS deployment will serve four subsidiaries under one customer contract, but legal requires strict data separation and the customer wants consolidated reporting. How do you design tenant isolation?
deploymentclouddesign
22
A financial customer prohibits public endpoints and needs AWS-to-Azure connectivity within 45 days. What private network design and proof would you propose?
designendpoints
23
Your platform, the customer, and a third-party model vendor each hold credentials for a 60-day deployment. How would you assign secrets and key-rotation ownership?
deploymentdependenciessecrets
24
An EU customer requires all personal data to stay in Germany, deletes source records after 30 days, and needs audit evidence for seven years. How do you reconcile retention and residency?
retentiondiscovery
25
During a 90-day deployment, the customer adds fields weekly to a Kafka and Snowflake contract consumed by five teams. How would you manage schema evolution?
schemasnowflakekafka
26
A billing integration may redeliver Kafka events after reconnects, but the customer asks for exactly-once invoicing before launch in six weeks. What do you promise and build?
kafkapromises
27
A 200-million-record migration passes row counts, but finance needs proof that monetary values and deletes are correct before a weekend cutover. What reconciliation do you run?
reactmigrations
28
Your AI workflow will handle 8,000 requests per minute across customer APIs, a queue, and a model provider. How do you define shared observability and SLOs?
sloobservabilityapi
29
Before a 24/7 production launch, the customer NOC, your support team, and the FDE pod use different severity models. How would you design customer-aware incident routing without discussing a specific incident?
incidentsdesignseverity-priority
30
A regulated customer has dev, validation, and production environments, with a weekly change window and no direct production access for your team. How do you promote releases?
validation
31
Two customer regions have drifted after engineers manually changed network rules and model settings during a 10-week rollout. How would you move infrastructure and configuration to code?
configiac
32
Thirty days before handoff, customer operations must meet a 45-minute recovery target for a Snowflake-to-AI workflow but has never run an index replay or model-provider failover. How do you prove operational acceptance?
indexessnowflake
33
A new retrieval release may corrupt derived indexes for 80 million documents, while source data remains intact. What rollback and replay design do you prepare before cutover?
indexesdesignrollback
34
A customer forecasts 3,000 requests per second normally and a 9,000-request burst during quarterly close. How would you run capacity testing before a 60-day launch?
capacitytesting
35
A claims team wants an LLM to approve claims automatically in eight weeks, but decisions are regulated and errors can cost $50,000 each. How do you assess workflow fit?
36
A global engineering customer wants RAG over 25 million manuals in 12 languages within 90 days, with answers traceable to approved versions. How would you design it?
design
37
A legal customer gives you six weeks and 400 historical queries to prove an AI research assistant. How do you build the evaluation set and customer rubric?
queries
38
A customer wants AI-generated maintenance instructions shown to 8,000 technicians, and a wrong torque value could damage equipment. How do you control hallucination and human review?
39
A pharmaceutical customer prohibits training on its data and considers prompts, retrieved passages, and outputs confidential. What architecture and commitments do you make?
architecture
40
A support assistant must answer in under 2 seconds, achieve at least 90% rubric pass rate, and stay below $0.04 per request for 20 million monthly requests. How do you set the budget?
41
A government customer permits only an approved Azure-hosted model, while your standard workflow uses another vendor and launch is ten weeks away. How do you handle the constraint?
procurementsoft-skills
42
You own a $6M account with three live workflows, declining weekly usage, stable uptime, and a renewal review in 75 days. How would you define and use a customer health score?
43
A deployed planning tool meets its technical SLOs, but only 22 of 300 planners use it after four weeks and regional managers prefer spreadsheets. What change-management plan do you run?
deploymentspreadsheetsspread
44
A financial-services launch is due in 70 days, but procurement needs a DPA, security review, vendor setup, and current SOC 2 evidence. How do you manage the critical path and evidence coordination?
schedulingprocurementcommunication
45
An EU customer wants a 60-day AI pilot using employee data from France and Germany. How do you coordinate GDPR work with Legal without acting as counsel?
gdpr
46
A customer-specific document workflow is priced at $900K ARR, but projected model, storage, support, and FDE costs vary sharply with volume. How would you assess pricing and account cost?
pricing
47
A strategic customer needs a bulk export feature for an 11-week launch, but product has no committed roadmap slot. How do you decide between a product gap and a local workaround?
roadmap
48
Three strategic customers request similar approval workflows, and your current account needs one in 90 days. What product feedback and RFC do you send internally?
feedbackdecision-making
49
A customer platform team has allocated only 240 hours to your 10-week deployment, but requests suggest more than 500 hours of meetings, data work, and integration support. How do you protect their engineering time?
deployment
50
At day 45 of a 90-day $5M deployment, the launch forecast slips by three weeks because private connectivity and data approval were underestimated. What executive status memo and decision log update do you issue?
deployment
51
A Tier-1 customer's 30-day PoC is blocked on legacy SAP, and the outcome controls a $400K deal; you must decide by day 8 whether to build an adapter or narrow scope. What do you do?
52
On day 10 of a 28-day analytics PoC, only 62% of required source records are complete, and success requires 90% coverage by day 21. How do you respond?
coverage
53
A customer's Snowflake SECURITYADMIN denies the service role 72 hours before a $650K deployment readiness review; the pipeline needs read access to 4 schemas. What is your plan?
deploymenthealth-checksci-cd
54
A Databricks source changed 7 of 46 columns overnight, breaking a Tier-1 pipeline 5 days before cutover; the customer will freeze schemas in 48 hours. How do you recover?
schemaci-cd
55
Kafka consumer lag reaches 18 million events during a customer pilot, alerts must be under 10 minutes, and the executive demo is in 36 hours. How do you triage it?
kafkaalerting
56
After a firmware update, 1,200 of 15,000 MQTT sensors report timestamps 20 to 45 minutes in the future, causing false anomaly alerts before Friday's factory-pilot review. How do you contain it and set the event-time policy?
alerting
57
A partner REST API allows 600 requests per minute, your initial sync needs 1.2 million records in 4 days, and cutover is next Monday. How do you stay within the limit?
rest
58
A nightly SFTP file arrives 9 hours late for the third time in 14 days, delaying a $300K customer's morning workflow; you have 48 hours to stabilize it. What do you do?
59
Workday and Salesforce disagree on employee IDs for 8% of 42,000 users, and access provisioning must launch in 7 days. How do you resolve identity ownership?
conflictownership
60
A SCIM integration missed deprovisioning 14 terminated users for 36 hours at a regulated customer, and the CISO expects a containment update in 2 hours. What do you do?
61
An SSO cutover locks out 22% of 1,800 pilot users, including the executive sponsor, and you must decide within 20 minutes whether to roll back. What do you do?
sponsorrollback
62
The customer's private link approval is delayed by 12 business days, but a 21-day PoC and $500K expansion decision end next Friday. What path do you propose?
63
A customer-facing mTLS certificate expired 3 hours before a 10-site rollout, and the maintenance window closes in 90 minutes. How do you recover?
mtls
64
A deployment log exposes 2 customer API secrets for 47 minutes, and the account's incident contract requires notification within 4 hours. What do you do?
incidentsapisecrets
65
A customer tester sees 3 records that may belong to another tenant during a $1.1M acceptance test, and the steering call starts in 60 minutes. What do you do?
acceptance
66
A 90-day backfill produces 6.4% different totals from the real-time stream, and finance must approve the customer rollout in 3 days. How do you reconcile them?
backfill
67
Duplicate events create 1,240 extra customer cases in 6 hours, and operations needs a safe replay decision before the 2 p.m. shift. What do you do?
68
A timezone bug leaves a 2.7% gap between the customer's daily ledger and your dashboard at month-end, and the CFO needs a signed reconciliation by 10 a.m. tomorrow. How do you handle it?
soft-skillsreact
69
A customer analyst's SQL scan is projected to consume 38% of the shared Snowflake warehouse during a 2-hour executive workshop. You have 30 minutes to decide whether to run it. What do you do?
sqlwarehousesnowflake
70
A Python adapter grows from 1.5 GB to 9 GB and crashes after processing 18 million rows; the customer's migration window ends in 6 hours. How do you recover?
concurrencypythonmigrations
71
Async retries raise a partner outage from 400 to 4,800 requests per second, and 6 customer deployments are affected; you must contain it within 15 minutes. What do you do?
deploymentasync
72
During a $900K phased cutover, aggregate error rate has stayed at 7.8% for 6 minutes against a 5% for 5-minute rollback gate, but one new region accounts for 94% of failures. What do you do?
aggregationrollback
73
A rollback restores application version 12, but 31% of customer records remain in version 13's partial state; payroll opens in 4 hours. How do you recover?
rollback
74
A bank asks you to deploy without logs, traces, or vendor metrics because of data policy, but go-live is in 10 days and the contract requires a 2-hour incident response. What do you propose?
deploymentmonitoringincidents
75
Your team proposes 99.5% availability, the customer demands 99.95%, and the $750K order form must be finalized in 3 business days. How do you resolve the SLO disagreement?
conflictformsslo
76
A security review with 86 controls is 2 weeks late, blocking a $1.3M production launch scheduled in 9 days. How do you move it without bypassing review?
77
Procurement wants go-live in 18 days, but private connectivity, pen testing, and data migration require 32 days; the $600K budget expires this quarter. What do you recommend?
procurementmigrationstesting
78
A customer's auditor requests 12 months of access-review evidence, but your package has a 7-week gap; renewal is in 14 days. What do you do?
79
Legal objects to a customer's request to retain EU prompts for 365 days, while the customer requires 180 days for investigations; contract redlines close in 5 days. How do you resolve it?
80
An LLM invents policy citations in 6 of 80 acceptance tests, and a $450K support rollout is due in 7 days. What do you do?
acceptance
81
Retrieval misses the correct contract clause in 18% of 250 tests, and the legal-operations pilot decision is next Tuesday. How do you improve it?
testing
82
An LLM eval reports 92% overall accuracy, but the customer's high-value claims slice is only 61% across 180 cases; launch approval is in 4 days. What do you recommend?
83
A red-team test gets the deployed assistant to reveal hidden instructions in 3 of 120 prompt-injection attempts, and production starts in 72 hours. What do you do?
injectiondeployment
84
A model vendor may retain PII from 4 customer fields for 30 days, but the healthcare pilot starts in 6 business days. How do you handle the vendor issue?
procurementsoft-skillspii
85
LLM cost per completed case rises from $0.42 to $0.88 after a prompt change, and the $700K account must approve unit economics by Friday. What do you do?
86
LLM p95 latency grows from 4 seconds to 19 seconds, causing 27% of agents to abandon the workflow; a 500-seat rollout decision is in 5 days. What do you change?
latency
87
A customer routes AI reviews between a day team that clears 500 cases and a night team that clears 700, but 18% of handoffs lose ownership; 900 new cases arrive daily, the backlog is 4,200, and the 8-hour promise is already missed. What do you change within 24 hours?
backlogpromisesownership
88
The deployed assistant scores 94% on quality, but only 11% of 640 eligible users adopt it after 30 days; renewal evidence is due in 2 weeks. What do you do?
decision-makingdeployment
89
Your health score says 82 out of 100, but the sponsor missed 3 meetings and weekly active use fell 38%; the quarterly account review is in 48 hours. How do you present the account?
sponsor
90
A proposed expansion adds $480K ARR but requires 1,600 engineering hours in the next 6 months; the investment decision is Friday. How do you evaluate it?
decision-making
91
A customer asks you to write directly to production Oracle tables because its supported API handles 2,000 records per day but 50,000 records must be loaded before cutover in 5 days. What path do you choose?
api
92
Your platform will retire connector API v1 in 90 days, but a customer's change freeze starts in 30 days and lasts 4 months; v2 changes authentication and pagination. How do you protect the live account?
paginationauth
93
Sales promised a 14-day production deployment to a $900K prospect, but technical discovery shows a minimum of 6 weeks; the kickoff is tomorrow. What do you do?
initiationdeploymentpromises
94
Two customer business units evaluate the same PoC: operations requires a 30% cycle-time reduction, while risk requires critical omissions below 1%; the purchase committee meets in 3 weeks. How do you settle acceptance?
decision-making
95
The executive sponsor for a $1.5M deployment leaves 10 days before go-live, and the replacement has not accepted the success criteria. How do you proceed?
fundamentalsdeploymentsponsor
96
The customer's 3 integration engineers are unavailable for 2 weeks during a 30-day PoC, and their work controls 5 required interfaces. What do you do by tomorrow?
types
97
A PoC has run for 11 weeks, consumed 720 engineering hours, has no active sponsor, and shows only $80K potential ARR; you must decide by Friday whether to kill it. What do you do?
sponsor
98
A strategic $2.4M account has missed 3 milestones, requires 900 more engineering hours, and renews in 45 days; leadership wants a save-or-stop recommendation in 72 hours. What do you do?
milestones
99
A customer deployment incident caused 2 hours of downtime and 1,850 delayed transactions; the customer expects a postmortem in 5 business days. How do you own it?
incidentstransactionsdeployment
100
A $1.8M renewal executive readout is in 4 days; deployment time improved 35%, but adoption is 24% below target and 2 critical integrations remain late. How do you present it?
decision-makingdeployment

Questions

1You inherit a $3M strategic account with a 12-week deployment window, six customer teams, and no agreed technical sponsor. How would you score discovery and map stakeholders in the first five days?

2A retailer says it wants an AI assistant to make store operations better within 90 days, but cannot define better. How do you turn that objective into a measurable workflow?

3A bank wants to automate a 14-person loan-review process in ten weeks, but has no reliable process metrics. How would you establish the current-state workflow and human baseline?

4A manufacturer expects $5M annual value from a predictive-maintenance deployment, while its platform team estimates 1,200 engineering hours to integrate it. How would you build the ROI map?

5You are the primary technical owner for a $4M account that must move from discovery to production in 90 days across Snowflake, Okta, and two internal APIs. What does your 30/60/90-day plan contain?

6A healthcare customer wants a production AI workflow in eight weeks, but identity access takes ten days and representative data has not cleared privacy review. How would you run an integration-readiness review?

7A 45-day fraud-detection PoC for a $2.5M account needs a custom data adapter and customer labels that may arrive late. What kill criteria would you agree before coding?

8A 70-day deployment depends on the customer's data, IAM, network, legal, and operations teams plus your product team. How would you define dependencies and RACI without creating meeting overhead?

9A logistics customer requests a routing workflow in ten weeks; 70% fits your standard template, 20% needs a supported extension, and 10% appears customer-specific. How do you choose what to build?

10At week six of a 12-week deployment, the customer COO adds two regions and a Workday integration while keeping the original launch date. How do you control scope?

11A fintech stores 12 TB in Snowflake and wants a customer-support AI workflow live in 60 days with source-to-answer freshness under ten minutes. How would you design the pipeline?

12A media customer has a Databricks lakehouse with 800 million documents and expects a six-week AI search pilot. What integration would you propose?

13A marketplace will send 60,000 order events per second through Kafka to your platform, and three customer teams consume the results. What event contract do you establish?

14A mining customer has 15,000 edge sensors sending MQTT over unreliable links and wants anomaly detection within two minutes. How would you design the telemetry path?

15A retailer updates 40 million product rows nightly but wants price changes reflected within five minutes before a 75-day launch. Would you use CDC or batch?

16A regional insurer offers a REST API, nightly encrypted files, and an internal queue for the same claims data, with a 12-week deployment deadline. How do you choose the integration interface?

17You must migrate three years of support tickets and then ingest live updates without downtime in a 50-day cutover. How do you combine backfill and incremental processing?

18A lending workflow depends on 25 customer fields, but null rates and code meanings vary by region and the launch is nine weeks away. What data-quality contract do you create?

19A 30,000-user enterprise requires SAML SSO, SCIM provisioning, and contractor separation before an 80-day launch. How would you design identity integration?

20A bank has 18 internal roles and your product has six permission bundles; go-live is in seven weeks. How do you map RBAC without creating one-off roles?

Locked questions