Influencer campaign data commonly arrives with mixed currencies, percentage symbols, compact follower units, missing values and implausible observations. Forecasting directly from those inputs produces fragile decisions precisely when marketing teams need confidence.
Marketing AI · Public case study
InfluenceLift AI
An end-to-end marketing analytics platform that turns unreliable influencer-campaign data into defensible sales forecasts, scenario comparisons and decision-ready outputs.
- Product surface
- Streamlit application
- Prediction service
- FastAPI
- Selected model
- Regularised Ridge regression
- Public proof
- Live demo, CI and model card
01 · Problem
Why this system exists.
Messy campaign input → schema validation → campaign cleaning → imputation and scaling → Ridge model → sales forecast and prediction interval → Streamlit simulator or FastAPI response
02 · Capabilities
What the public implementation demonstrates.
- Parses values such as currency-formatted spend, percentages and K/M follower notation.
- Audits missing values, invalid ranges, scale anomalies and schema problems before prediction.
- Produces campaign-level forecasts, empirical prediction intervals and extrapolation warnings.
- Supports interactive scenario comparison, batch CSV uploads and downloadable predictions.
- Runs linting, tests and deployment smoke checks through GitHub Actions.
03 · Ownership
What I designed and built.
- Problem framing and business-use-case definition.
- Reusable data-cleaning and model pipeline design.
- Model evaluation, selection and responsible interpretation.
- Streamlit product interface and FastAPI prediction service.
- Testing, CI, documentation, model card and deployment workflow.
04 · Engineering
Reliability, safety and responsible use.
Engineering decisions
- The public demonstration uses deterministic synthetic data because the original challenge dataset is not redistributed.
- The model communicates uncertainty rather than presenting forecasts as guarantees.
- The implementation separates reusable modelling code, API logic, interface code, tests and documentation.
Limitations
- Validated metrics describe the supplied case-study data and are not universal performance benchmarks.
- The model estimates associations, not causal effects; changing spend does not automatically cause the predicted sales movement.
- Production use would require organisation-specific data governance, monitoring and access controls.
05 · Technology
The implementation stack.
Public and inspectable.
The repository contains the implementation, documentation and setup guidance available for this project.