The problem with “AI projects” today
Most AI projects fail long before the first model is trained. Teams rush to algorithms before they understand the business outcome they need to improve, or what their data truly represents.
Across logistics, finance, healthcare and retail, the remedy has been consistent: real AI does not start with data. It starts with business intent and the mathematics that connects intent to intelligence.
When data is not enough
The dataset did not contain labels tied to the business metric that needed improvement. There was no obvious target variable. Before modelling, we worked with domain experts to define what the business was optimizing, why existing heuristics were failing, and which performance indicators actually defined success.
The warehouse data was real-world data: structured tables intertwined with nested JSON, missing values, inconsistent entries and outdated schema versions. We identified predictable operational patterns, invalid or extreme outliers, representation imbalances, and gaps such as missing timestamps or categories.
The toughest challenge: missing fairness
The hardest problem was not noise. It was representation bias. Some entities had rich historical data; others were new or underrepresented. A naive ML model would rank the well-documented entities higher and unfairly penalize newer entrants.
Fairness therefore had to be designed directly into the system - mathematically.
Translating business into math
The system needed more than a prediction. It needed to rank and recommend based on reliability, cost-effectiveness and performance while respecting business constraints. We developed a hybrid scoring equation that blended data-driven learning with rule-based fairness and policy alignment.
| Symbol | Meaning |
|---|---|
| p | Price or cost metric - lower values yield higher preference in cost-sensitive contexts |
| r | Rating or quality metric - reliability or satisfaction |
| f | Fulfillment or completion ratio - delivery consistency |
| a | Acceptance or responsiveness rate - engagement or willingness |
| h | Historical volume or experience - operational maturity |
| d | Distance or proximity factor - spatial or temporal efficiency |
| α, β | Weight coefficients balancing handcrafted fairness with learned intelligence |
The handcrafted score used tunable weights to reflect business priorities.
This let the system prioritize reliability and fulfillment over proximity when quality mattered more than distance.
The model that learned like a human
We began with gradient boosting using CatBoost, a strong algorithm for structured data. It performed well on known correlations, but could not capture contextual interactions among cost, distance and quality.
We therefore engineered a custom attention-based neural model in PyTorch. One sub-network encoded task-specific attributes, another encoded participant or provider behavior patterns, and the attention layer learned how those two dimensions interact.
This architecture allowed the system to reason about compatibility, not just similarity.
Proving it worked
We validated the system with live A/B testing against the baseline recommendation logic. The hybrid approach did not just perform better - it behaved better: explainable, ethical and business-aware.
| Metric | Before | After hybrid model |
|---|---|---|
| Engagement rate | Baseline | ↑ 25% |
| Completion rate | Baseline | ↑ 18% |
| Operational efficiency | Baseline | ↑ 15% |
| Fairness for new entrants | Low | High and balanced |
| Cost optimization | Uncontrolled | Strategically tuned |
The breakthrough was recognizing that pure ML predictions cannot independently respect fairness, cost boundaries and compliance rules. Those constraints must be embedded in the scoring logic.
The enterprise takeaway
- Start with the business objective, not the dataset.
- Translate goals into mathematical formulations, not just code.
- Build hybrid systems that combine data intelligence and policy logic.
- Make fairness, explainability and optimization part of the equation.
- Validate with real-world metrics, not only offline validation accuracy.
Frequently asked questions
Why does AI work start with business intent?
The business objective defines what should be optimized and which data, metrics and constraints are meaningful.
Why was a hybrid score used?
It balanced learned predictions with interpretable business rules, fairness considerations and operational constraints.
What did live validation show?
The hybrid approach increased engagement by 25%, completion by 18% and operational efficiency by 15%, while improving fairness for new entrants.
Originally published by Prestanda Consulting on Medium. View the original publication.