Azul Airlines AI Price Optimization

The core architectural decision was to build around flight-level analysis from the start, rather than adapting a segment-based system. Each of the 8 AI models generates recommendations for specific flights based on that flight's individual booking curve, competitive positioning, and demand signals.

On a single day like December 10th, the system adjusted prices for 1,306 flights, each with individualized recommendations. The Inventory Shaping system proactively manages 50,000+ flights with demand-based tags, enabling anticipation of inventory needs across Azul's entire network.

This granularity is what separates the system from traditional revenue management. Instead of applying a segment-wide price increase that helps some flights and hurts others, each flight is optimized independently. Opportunities invisible in aggregated data become actionable at the individual flight level.

Rather than building one general-purpose pricing model, we built 8 specialized models each addressing a specific revenue scenario.

For early booking, Stickler Pricing maintains price consistency across the booking journey, preventing trust-eroding fluctuations while allowing strategic adjustments. Journey Curve analysis optimizes how prices evolve from the initial booking window to departure, learning optimal curves from historical performance.

In the mid-cycle, the Inventory Shaping system proactively tags flights based on predicted demand patterns, enabling proactive rather than reactive optimization. A stateful architecture maintains each flight's demand profile as new booking data arrives.

For late-stage scenarios, Holiday Radar identifies peak travel period flights and applies specialized pricing logic. Dedicated models handle sold-out flight overbooking management and near-empty flight last-minute revenue capture, each with distinct optimization logic suited to its specific scenario.

Specialized models outperform general-purpose AI in complex domains. Each revenue scenario has distinct characteristics that require different algorithms, different data features, and different business rules. A single model trying to handle everything produces mediocrity across the board.

With 8 models running simultaneously, multiple models frequently generate recommendations for the same flight. A flight might qualify for both Holiday Radar pricing and Inventory Shaping adjustments at the same time.

We built a priority-based recommendation system that resolves these conflicts deterministically. Each model has a priority ranking based on the flight's current state and time to departure. The system selects the highest-priority recommendation, ensuring each flight receives exactly one optimized price adjustment.

This architecture prevents recommendation chaos while preserving the specialized logic each model provides. The priority system also includes a confidence scoring mechanism that flags low-confidence decisions for human review, improving pricing accuracy by 23% compared to single-model recommendations.

We built FCC Live, a real-time execution engine that runs models continuously throughout the day. The system uses Python with Celery task queues to orchestrate model execution, with Snowflake serving as the data warehouse processing approximately 20 data tables daily covering booking information, competitive pricing, and historical performance.

Each model run pulls the latest booking and market data from Snowflake, generates flight-specific recommendations through the specialized models, applies priority resolution for conflicting recommendations, and outputs pricing adjustments to the automated workflow. The entire pipeline runs without human intervention.

The data warehouse was structured specifically for efficient flight-level queries across 50,000+ flights, with careful indexing and partitioning strategies to maintain query times under 200ms even during peak processing.

The breakthrough that enabled real-time pricing was eliminating the human approval bottleneck. We built automated FTP integration with Azul's Navter pricing system. The AI generates recommendations, the system validates them against business rules, and approved changes transfer directly to Navter without manual intervention.

This reduced pricing update frequency from once daily to every 50 minutes, a 28x increase in update frequency. The system can now respond to competitor price changes, sudden demand spikes, and booking velocity shifts within the same hour.

Automation at this level requires trust built through validation layers, not perfect models. Multiple validation layers were built in: business rule constraints that prevent recommendations outside acceptable ranges, anomaly detection that flags unusual recommendations for human review, a circuit breaker mechanism that reverts to manual pricing if data quality issues or model degradation are detected, and a gradual rollout that tested models on subsets of flights before full deployment.

Real-time systems that run continuously without persistent state produce contradictory recommendations over time. The system maintains historical context across every model run, tracking what recommendations were made, what was implemented, and how flights performed.

Snowflake stores this historical state while Celery manages task execution and workflow orchestration. This combination provides the data persistence and workflow coordination needed for continuous operation.

The stateful architecture enables learning without retraining from scratch. The Inventory Shaping model improved forecast accuracy by 31% compared to Azul's previous manual approach by continuously adjusting forecasts when actual booking patterns deviated from predictions. Models refined their strategies based on accumulated production data, which is reflected in the January 2026 acceleration: the system gets better as it runs.