AeroPulse

Operational issues can become visible only after they have already created an impact.

Aviation operations generate large amounts of information across flights, airports, aircraft, routes, turnaround activity, passenger loads, maintenance indicators, operational events and disruption conditions.

The challenge is not simply having data. The challenge is turning that data into signals that operations teams can investigate early.

Designed around operational decision-makers.

From operational data to actionable signals.

AeroPulse transforms flight-operation records into a layered intelligence system rather than presenting users with another static reporting dashboard.

The MVP focuses on six intelligence layers.

Building a realistic synthetic aviation dataset.

Because operational aviation data is sensitive and difficult to access for experimentation, AeroPulse uses synthetic data designed to represent common aviation operational dimensions.

Key data dimensions

• Flight schedules and actual operation times
• Airlines and aircraft types
• Origin and destination airports
• Turnaround duration
• Passenger capacity and passenger volume
• Aircraft utilisation
• Weather disruption indicators
• Airport congestion indicators
• Maintenance indicators
• Operational events
• Departure and arrival delays

Product decisions started with analytics, not machine learning.

Before introducing predictive modelling, I created an analytical layer to understand the operational dataset.

This included overall KPIs, airport performance, route performance, aircraft performance, airline performance, event-level analysis, daily performance, operational pressure and delay severity.

Finding unusual operational patterns.

AeroPulse uses Isolation Forest to identify records whose operational characteristics differ from the broader dataset.

The model considers variables such as delay, turnaround variance, load factor, aircraft utilisation, congestion, weather, maintenance and operational pressure.

Predicting potential delay risk.

The predictive layer uses Logistic Regression to estimate the probability that a flight will be delayed.

Categorical variables are encoded using one-hot encoding, while numerical features are imputed and standardised through a preprocessing pipeline.

Why these metrics matter

Because the synthetic dataset contains more delayed than non-delayed flights, accuracy alone does not fully describe model performance. Precision, recall, F1 and ROC-AUC provide a broader view of how effectively the model separates delayed and non-delayed records.

Combining multiple signals into one monitoring layer.

The operational risk engine combines predictive delay probability with anomaly signals, operational pressure and operational conditions.

Delay Risk                  50%
      │
      ├──────────────┐
      │              │
Anomaly Signal       │      25%
      │              │
      ├──────────────┤
      │              │
Operational Pressure │      15%
      │              │
      ├──────────────┤
      │              │
Operational Conditions      10%
      │
      ▼
Operational Risk Score
      │
      ├── LOW
      ├── MEDIUM
      └── HIGH

The resulting score is intended to help prioritise operational investigation. It is not a certified aviation safety prediction.

Turning data into an executive-friendly narrative.

AeroPulse uses a Groq-powered language model to analyse the aggregated operational outputs and generate concise insights for decision-makers.

The AI layer is deliberately constrained to the supplied dataset. It is instructed not to invent causes, statistics, airports, routes or operational facts.

From raw operational data to decision support.

                 RAW AVIATION DATA
                        │
                        ▼
                DATA GENERATION
                        │
                        ▼
              FEATURE ENGINEERING
                        │
                        ▼
              OPERATIONAL ANALYTICS
                        │
             ┌──────────┴──────────┐
             ▼                     ▼
      ANOMALY DETECTION      PREDICTIVE MODEL
             │                     │
             └──────────┬──────────┘
                        ▼
                OPERATIONAL RISK
                        │
                        ▼
                AI INSIGHT LAYER
                        │
                        ▼
                STREAMLIT DASHBOARD
                        │
                        ▼
               HUMAN INVESTIGATION

A complete data-to-product workflow.

• Defined the aviation operations problem and product question.
• Defined primary and secondary users.
• Generated a 10,000-record synthetic aviation dataset.
• Built a reusable feature-engineering pipeline.
• Created operational KPI and performance analytics.
• Implemented Isolation Forest anomaly detection.
• Built a Logistic Regression predictive delay-risk model.
• Created a multi-signal operational risk engine.
• Integrated a grounded AI-assisted insight layer.
• Built an interactive Streamlit decision-support dashboard.
• Documented responsible-AI limitations and intended use.
• Deployed the interactive prototype for public demonstration.

Designed as decision support, not autonomous control.

AeroPulse is an experimental portfolio prototype using synthetic data. Its outputs should not be interpreted as real-world aviation safety conclusions.

• The dataset is synthetic.
• Model relationships are illustrative.
• The system does not establish causality.
• Operational risk scores are prioritisation signals.
• AI-generated insights require human review.
• The system is not designed for autonomous aircraft, equipment or operational control.
• The prototype is not intended for regulatory certification or safety-critical deployment.

Technology stack.

From prototype to operational intelligence platform.

The next iteration would move AeroPulse from a synthetic demonstration toward a more realistic operational intelligence system.

• Integrate validated historical aviation datasets.
• Add real-time operational event streams.
• Add richer weather and airport congestion data.
• Introduce real-time alerts and threshold monitoring.
• Add deeper root-cause investigation workflows.
• Validate model performance against real historical outcomes.
• Add role-based dashboards for different operational teams.
• Expand the API layer for integration with operational systems.