01 — Research problem
International comparison fragments when every market uses its own format.
Local studies produce rich information, but formats, languages and levels of synthesis vary. An insight from Poland may be phrased, categorized and archived differently from a signal observed in Portugal. The central team then spends substantial time harmonizing deliverables before comparing findings.
I framed the problem as a reproducible transformation chain. The goal was to reduce the mechanical work of transcription, translation, rewriting and classification while retaining the source, market, severity and context of each observation.
02 — Design choice
Build a working tool to test the contract between the team and artificial intelligence.
The prototype accepts three entry points: verbatim, audio and public reviews. Audio is transcribed. The application analysis service processes original verbatim without prior translation; the command-line pipeline separately provides language detection, translation and glossary support. Free or guided extraction structures responses into shared fields.
The output follows a strict JSON — JavaScript Object Notation — schema imposing the same fields on every response: theme, severity, insight, recommendation, provenance, market and potential cross-market signal.
03 — Research product
The repository turns extractions into usable team memory.
Each insight can be reviewed, edited, filtered by market, source or theme and linked to its original data. A question and group library supports reusable interview guides. A control dashboard aggregates studies and surfaces time-to-insight as well as recurring themes.
The system covers four markets, four source connectors, two analysis modes and four main themes. Most importantly, the prototype shows that a team can share one method without forcing every country into the same collection channel.
04 — Scope
Automation prepares comparison; research validation remains essential.
The model prepares an editable first pass. Reviewing wording, severity and recommendations remains necessary before use; the prototype does not enforce an approval gate before distribution. Time savings and review quality are evaluation targets.
For a ResearchOps team, the next evaluation should measure inter-researcher agreement, actual time saved, translation quality by market and the proportion of corrected extractions. These indicators evaluate the system as a research tool rather than a standalone artificial-intelligence demo.
05 — From concept to decision
The same schema does not guarantee the same meaning.
Consider two comments: “I do not know what happens after confirmation” and “I would like to undo this.” They may share a theme while suggesting different actions: explaining what follows versus providing a reversible control.
To evaluate the prototype, I would compare extractions with researchers’ work, retain disagreements and measure review time. The system’s value would be the quality of the corrected synthesis, with provenance that is easy to retrieve.