Data Analyst

Nummo is hiring a Data Analyst. This posting is invented. There is no opening, no recruiter, and no spreadsheet of candidates waiting behind it. The description exists because someone asked for one, which is also how a lot of real analytics work begins. You will answer questions that sound small and are not. “How many people used this last week?” depends on what “used” means, which week, which accounts, and whether a test user counts. You will ask those questions before you write the query. If you skip them, you will produce a number that is precise and wrong, and someone will paste it into a slide. Getting the number back out of the slide is harder than writing the query. The data lives in a database that was shaped by the product, not by your report. Tables have names that made sense to the person who created them. Some events are missing because the client never sent them. Some rows are duplicated because a job ran twice and nobody noticed until the chart spiked. You will learn which tables lie, which ones are merely incomplete, and which ones are fine if you filter out the internal accounts. You will write that down somewhere other than your head, because the next person to ask is you, in three weeks, with no memory of the filter. A normal request arrives as a sentence. “Can we see if the new flow is working?” You will turn that into a definition: who entered the flow, who finished it, in what time window, excluding employees. You will show the count, the rate, and the number of accounts behind the rate. A rate with twelve accounts is a story about twelve accounts. You will say that in the same message as the chart, so the chart cannot travel alone. You will use SQL. You will also use a notebook or a dashboard tool when a person needs to look at the result without asking you again. The tool matters less than the habit of tying every figure to a query someone else can rerun. If the number came from a spreadsheet you edited by hand, the number does not leave your machine until you can explain the edits. Hidden edits are how a metric starts meaning two things. You will spend time on definitions. “Active” is not a column. It is an agreement. Once the agreement exists, you will keep using it, even when a different definition makes the chart look better. Changing a definition is allowed. Changing it in a footnote after the meeting is not. When product, sales, and engineering each mean something else by the same word, you will pick one meaning, name it, and put the other meanings in a second table instead of averaging them into a third meaning that nobody wanted. You will be asked for predictions. Sometimes the honest answer is a range. Sometimes the honest answer is that the history is too short, or the change is too new, or the accounts are too different to pool. You should be willing to say that. A confident guess with no base rate is worse than a late answer. You should also be willing to give the simple answer when the simple answer is true. Not every question needs a model. Some questions need a count and a date. Privacy is part of the work. You will look at records that belong to customers. You will aggregate them before they go into a deck. You will not export a list of people to settle a curiosity, and you will not drop raw rows into a chat because a chart was annoying to make. If a result would identify a small group, you will widen the group or drop the cut. The point of the report is a decision, not a tour of someone else’s data.