Sign in with your Google account to start the investigation. Only invited candidates can access this test.
What Trigo does: Trigo powers frictionless checkout in retail stores: cameras and AI watch what customers take from the shelf and charge them automatically, without scanning each item. The system must be accurate—customers pay the right amount, and the store catches theft.
Why human tagging: When the AI is unsure (e.g. unclear view, overlapping items), the transaction is sent for manual review. A human annotator watches the same footage and tags the outcome—for example “theft” or “not theft,” or corrects the list of items. This tagging does two things: it fixes the charge for that transaction, and it helps improve the AI over time. So the data you see mixes AI-generated logs (what the system thought) and human review outcomes (what was finally decided).
We have given you a dataset of transactions, AI logs, and human tagging. Your job is to go beyond the numbers: find where the system struggles, where theft might be slipping through, and how the patterns relate to real store behavior. Use SQL to investigate, then interpret and report what you find.
You'll work through the case in 3 phases:
Use Start investigation above to open the test — that starts your 90-minute timer. The Test tab is locked until then.
stores
id — store idname — store namecurrency — currency codereceipts
storeId — store referencereceiptId — receipt idtimestamp — transaction timescoId — self-checkout iditems — item count or listtotalValue — total valuetasks
taskId — task idreceiptId — receipt referencetagStartTimestamp — review start timetagEndTimestamp — review end timetag — 'theft' or 'not theft'details — optional notesannotatedBy — reviewer idThe last task per receipt (by tagEndTimestamp DESC) is the final decision.
If you run out of time, write how you would have approached the remaining steps.
Run a simple query to verify the database is reachable and see sample data (and data types). Set the Backend URL in the Submit tab if needed.
Check = run your query against the evidence. Submit answer = lock in your finding. First try correct = better.
B2 — You found a checkout with a high theft rate. But is it real, or is the data misleading you? Give two possible explanations — one data issue, one real-world reason. Write 2–3 sentences.
D — The 40% alarm: someone reports "Store X has 40% theft." The store manager says it can't be right. In 4–6 sentences: what could be wrong, one query you'd run to check, and one sentence you'd tell the manager right now.
C2 — In production, this table has millions of rows. In 2–3 sentences: when does partitioning and clustering make the biggest difference? What goes wrong if someone queries without filtering by date?
Sign in with Google to submit your investigation. We keep your most recent submission.
Set this to your Cloud Run URL or http://localhost:8080 when running locally.
Sign in with Google above to submit.