Project Manager

About the task:

We monetize SMS traffic. The economics are calculated to the penny. We need a person who handles this loop daily: guards the economics, runs experiments instead of arguments, and doesn't let a decision be made on a figure that hasn't yet matured.

What to do:

● Manage the hypothesis queue. Formulate, estimate in rubles before launch, prioritize, and bring to the stage of "implemented" or "closed." Hypotheses don't come from above – you find them yourself by looking at the data.

● Bring the issue to a number and be responsible for that number. Precisely formulate what to calculate and on what sample, get the calculation – by your own hands or others' – and verify it before making a decision. The cost of an unverified number here is millions per month.

● Design A/B tests. Before launch: metric, sample size, duration, stopping criterion, what we consider success. Our conversion rate is 0.15% from delivery – experiments need to be calculated for rare events.

● Read results honestly. Adjust for conversion maturation, separate signal from noise, catch false positives. The ability to say "not enough data yet" is valued more than the ability to show growth.

● Be a bridge to development. Turn a hypothesis into a technical specification, discuss implementation cost, accept the result, verify that exactly what was intended has been rolled out.

● Maintain daily control. Sending cost, CTR, conversion to payment, average check – by operator, segment, and source. A tariff change or a conversion drop should be flagged the same day, not weeks later.

● Guard the methodology. Catch look-ahead bias, selection bias, survivorship bias, and Simpson's paradox – in others' calculations and your own. This is the main protection against costly errors.

● Calculate thresholds for negotiations. Break-even channel price, segment rejection threshold, update delay cost – so that negotiations with a supplier are based on numbers, not feelings.

Mandatory:

● Setting up calculations. Formulate the task so that the result can be verified: what period, what slice, how exactly the metric is calculated. A number without origin is not accepted.

● Verifying others' results. Cross-check the result with a known value, look at raw data, recalculate in another way. Know the common ways calculations lie: the wrong slice was taken, the metric definition changed along the way, zero was substituted for a null, data after the measurement was included in the feature.

● Data biases. Look-ahead bias, selection bias, survivorship bias, Simpson's paradox. The ability to build a feature strictly "as of a date" and explain why it's not the same as a date filter.

● Statistics for A/B tests. Power and sample size, confidence intervals, rare events, multiple comparisons. Understanding when zero conversions is a signal and when it's just a small number of observations.

● Unit economics. ROI, CPA, ₽ per number, payback threshold, channel profitability. The ability to calculate in half an hour at what price a channel stops paying for itself.

● Attribution and lags. Conversions come later than impressions – the ability to measure the maturation curve and adjust recent data accordingly.

● Setting tasks for development. Clear technical specifications, acceptance, understanding of the technical cost of a solution.

Will be a plus:

● SQL by hand: window functions, ASOF and interval joins, aggregations over hundreds of millions of rows – to verify others' calculations and argue with them faster

● ClickHouse specifically: partitions, FINAL for ReplacingMergeTree, join algorithms

● CPA, arbitrage, lead generation: postbacks, offers, storefronts, subs

● SMS and telecom: operators, alpha-names, tariffs, deliverability

● Python/pandas for tasks inconvenient in SQL

● Experience negotiating traffic prices with suppliers

● Your own metric monitoring tool that you've brought to a working state