AI Space Utilization Analyst Interview Questions

Great answers define utilization as actual occupied space divided by total available space over a time period, and discuss peak vs. average utilization.

Cover PIR, BLE beacons, WiFi probe requests, LiDAR, and camera-based systems - noting privacy, accuracy, cost, and granularity trade-offs.

Occupancy is presence; utilization is productive or intentional use of a space. A person sitting in a conference room alone differs from a fully-booked meeting.

A heatmap visualizes density or intensity of space usage across a floor plan, typically generated from sensor coordinates aggregated into grid cells.

Cover hybrid work, rising real estate costs, sustainability targets (carbon per sq ft), and employee experience optimization.

Address null handling, sensor drift calibration, duplicate deduplication, timestamp normalization across time zones, and outlier detection for erroneous readings.

Discuss feature engineering (cyclical encoding for day/time, holiday flags), model choices (Prophet, SARIMAX, gradient boosting), and cross-validation strategy for time-series.

Cover YOLO-based detection, tracking with DeepSORT or ByteTrack, handling occlusions, lighting changes, and privacy-preserving edge deployment.

Discuss Kalman filtering, Bayesian fusion, or ensemble averaging; address sensor confidence weighting and temporal alignment.

Analyze peak concurrent occupancy distributions, simulate desk-sharing schedules, assess department-level variation, and model employee friction impacts.

Cover cost-per-occupied-sq-ft reduction, subletting revenue potential, energy savings, employee productivity proxy metrics, and payback period modeling.

Discuss streaming data pipelines (Kafka, IoT rules engine), threshold-based alerts vs. ML anomaly alerts, and notification channels (Slack, email, digital signage).

Discuss on-edge processing (no image storage), anonymization techniques, DPIAs, legal basis for processing, and transparency to employees.

Spatial resolution = granularity (desk-level vs. floor-level); temporal resolution = sampling frequency (1-min vs. 15-min). Both affect model choice and dashboard design.

Check sensor uptime logs, correlate with HR data (headcount changes), compare against WiFi vs. badge vs. camera sources, and test for structural breaks statistically.

Cover streaming ingestion (Kafka/IoT Hub), 3D spatial modeling (Unity/Unreal + BIM), real-time ML inference, edge computing, and the feedback loop to building management systems.

Define state (current occupancy, upcoming bookings, employee preferences), action (assignment decisions), and reward (utilization efficiency + satisfaction); discuss simulation environments.

Model zones as nodes, transitions as weighted edges, use GNN (e.g., GraphSAGE) to learn spatial-temporal flow embeddings for predicting congestion and optimizing routing.

Map patient flow with discrete event simulation, identify spatial bottlenecks with queueing theory, use RL or optimization (linear programming) for layout scenarios, and validate with historical A/B data.

Use ensemble of statistical methods (isolation forest) and cross-sensor validation; if one sensor spikes but adjacent sensors don't corroborate, flag as malfunction.

Cover feature store, model registry (MLflow), CI/CD (GitHub Actions), monitoring for data drift and concept drift, automated retraining triggers, and A/B deployment with shadow mode.

Use BLE/WiFi co-location data to build interaction graphs, apply network analysis (clustering coefficient, betweenness centrality), and correlate with org-chart data.

Model warehouse as a grid graph, integrate forklift/AMR telemetry with occupancy data, use OR-Tools or custom RL for zone assignment and path optimization.

Discuss difference-in-differences, synthetic control methods, or randomized controlled trials with staggered rollout; address confounders like seasonality and team composition changes.

Discuss federated analytics, differential privacy, aggregated-only reporting per tenant, and technical data isolation architecture with role-based access control.

Analyze vacancy distribution (is it uniform or concentrated?), model consolidation scenarios with peak-occupancy headroom, compare costs (moving vs. redesign), and present a decision matrix.

Discuss domain adaptation, transfer learning, sensor calibration differences, cold-start strategies (use building metadata as features), and the need for a warm-up period.

Acknowledge the anomaly transparently, check sensor health status, cross-reference with badge/camera data, explain likely causes (sensor duplication, bad calibration), and commit to a post-meeting investigation.

Use multi-objective optimization (Pareto front), segment by store archetype, A/B test layout variants, and define a composite score weighting both metrics by business priority.

Present data alongside qualitative employee survey data, propose a phased pilot with measuring points, offer compromise (focus rooms + open zones), and frame it as an experiment not a mandate.

Start with badge data analysis (entry/exit patterns, peak times, floor-level inference), supplement with WiFi client counts, propose a sensor pilot on one floor, and build quick-win dashboards to justify investment.

Present honest findings, suggest schedule optimization and room-sharing strategies first, quantify deferred capital expenditure savings, and recommend data-driven phased expansion.

Investigate whether the model accounted for seasonal inventory spikes, check if downstream processes (packing, shipping) were bottlenecked, roll back if necessary, and update the model with the new data.

Use transfer learning from similar climates/mall types, incorporate calendar-based features (Ramadan, school holidays, weather), start with rule-based baselines, and iteratively improve as real data accumulates.

Define a canonical data schema, build an ETL normalization layer per site, implement data quality scoring, and create tiered dashboards (global overview + drill-down per region) with confidence indicators.

Cover chain design: data retrieval tool → summarization chain → style/formatting prompt → output parser; discuss handling of numerical accuracy and hallucination mitigation with retrieval grounding.

Discuss dataset creation from labeled sensor streams, model selection (TabNet or time-series transformer), training pipeline with HuggingFace Trainer API, and evaluation with confusion matrix analysis.

Cover model optimization (TensorRT quantization), containerized deployment with Docker, centralized model updates via OTA, monitoring with Grafana, and fallback strategies for device failures.

Define functions for date ranges, zones, metrics; build a prompt system that maps natural language to structured function calls; handle ambiguity with clarification prompts; ensure SQL injection safety.

Cover experiment naming conventions, metric logging (MAE, MAPE by time horizon), model registry with staging/production stages, and automated comparison dashboards.

Define the causal graph (policy → utilization, confounders: weather, season, team size), use propensity score matching or instrumental variables, and validate with placebo tests.

Cover Kafka topic partitioning strategy, consumer group design, exactly-once semantics, schema registry for sensor event formats, and integration with real-time feature store for ML inference.

Describe loading floor plan geometries, overlaying sensor point data, spatial joins to assign readings to zones, buffer analysis for sensor coverage gaps, and aggregation into zone-level time series.

Use collaborative filtering on booking history, graph-based proximity features (org chart + floor layout), constraint satisfaction for capacity, and a ranking model (LightGBM or neural ranker) for suggestions.

Use statistical tests (KS test, PSI) on feature distributions, maintain reference windows, set up automated alerts via Grafana or Evidently AI, and trigger retraining pipelines on drift detection.

Look for evidence of data-backed confidence, diplomatic framing, empathy for the leader's perspective, and a concrete outcome.

Strong answers show comfort with uncertainty, creative data augmentation, transparent communication of confidence levels, and iterative refinement.

Assess translation skills between technical and non-technical audiences, active listening, and the ability to find shared objectives.

Look for intellectual humility, root cause analysis skills, accountability, and concrete process improvements implemented afterward.

Assess for framework-based prioritization (impact vs. effort), stakeholder communication skills, and ability to set expectations diplomatically.