DFreight is a B2B digital freight forwarder. Its operations teams handled shipment information across emails, documents and internal workflows where inconsistent formats and important exceptions made automation difficult.
The first evidence came from a focused email extraction MVP. We then went further and designed V1 of an end-to-end customs declaration workflow, from extracting information across different sources to cross-checking it and drafting the form for submission.
What needed to change
The opportunity was not simply to add AI. The team needed to decide what to automate, how to reconcile conflicting or incomplete source data, where human review still mattered and how to evaluate each step when false positives and missed information carried different operational risks.
How I approached it
- Worked with Operations, Data and Engineering to map the customs-declaration workflow, identify high-value fields and translate operational needs into requirements and acceptance criteria.
- Scoped a focused email extraction MVP to establish technical feasibility before expanding the product surface.
- Designed V1 of the end-to-end system to extract information from different sources, normalize and cross-check the values, surface conflicts and draft a ready-to-submit customs declaration form.
- Defined evaluation priorities by field and workflow step, deciding when precision or recall mattered more based on the cost of false positives versus missed information.
- Kept human review, exception handling and the operator's final decision inside the product workflow rather than treating model output as the finished result.
The choices that shaped the work
Build evidence with a bounded extractor
Field-level extraction from email created a clear first test surface before the team committed to the wider customs-declaration workflow.
Design the workflow, not only the model
V1 connected extraction, normalization, cross-checking, conflict handling and form drafting so model output could support a real operational task.
Let failure cost choose the metric
I compared the consequences of false positives and false negatives for each field and step, then prioritized precision or recall accordingly instead of relying on one aggregate accuracy score.
What changed — and what the result means
The email extraction MVP achieved approximately 85% average accuracy per field, providing a concrete basis for deciding where the solution could support operations and where further work was needed.
The later V1 design extended the work into an end-to-end customs-declaration system, connecting multiple information sources, cross-checking and a drafted form while keeping the operator in control.
The evaluation framework gave the team a clearer way to make product trade-offs between precision, recall, automation coverage and review effort.
WHAT THIS REINFORCEDAI product work is not about maximizing one score. The right metric depends on the decision being supported, the failure cost and the point where human judgment still belongs in the workflow.