AI Intern @ Ginjer AI
Summer 2026
Ginjer AI is an AI-powered platform giving advertisers transparency into the paid media dynamics of an industry and its players.
In charge of data quality issues. Solve those issues by analysing, understanding and rebuilding of pipeline and workflows with multidependecies.
Unresolved ad volume, before & after
How I approached it
The brief was “fix the taxonomy.” The actual problem traced back further: several vocabularies were coexisting across the pipeline without ever being reconciled, and a value computed nightly by the labeling model wasn't being read by anything downstream. I measured the real error rates before touching any category, and caught three successive measurement biases in the process, each of which would have sent the fix in the wrong direction.
Redesigned the labeling schema to separate concepts that were being conflated (who is paying for an ad versus what the ad actually shows), and replaced a fully manual, spreadsheet-based matching process with an automated resolution pipeline: exact lookup first, model inference only for what the lookup couldn't resolve.
Consistently chose deterministic rules and closed enumerations over asking the model to “behave well.” Concretely: a dictionary lookup resolved most cases before any model call, ambiguous cases were decided by explicit boolean rules instead of a confidence score, and automated tests failed the build whenever a taxonomy and a prompt drifted apart.
Built two evaluation harnesses: one to compare successive prompt versions against a held-out test set, another to compare three competing pipeline architectures on the same real ads, measuring cost, latency, and reliability side by side. The final architecture was chosen because it was measurably cheaper for a near-identical quality, not because it seemed more robust.
Deployed the redesigned pipeline to production with a full historical backfill, added monitoring signals to catch silent regressions, and documented the remaining technical debt and open decisions for the team.