Analytics & Tracking · Mumbai (on-site) · Full-time

Data Scientist

Own the streaming pipelines and warehouse models behind Cityflo's live decisions (ETA, occupancy, lateness, fraud) and the judgment calls underneath them.

You'll own the data behind Cityflo's live decisions: the streaming, event-driven pipelines that power ETA, occupancy, dynamic routing, and fraud signals, and the warehouse models that analytics, ML, and product teams build on. You sit across data science and engineering: you design the pipeline and you reason about whether the number coming out of it is actually true. You'll work directly with product, ops, and business to turn open-ended questions into solutions that move revenue, margins, and fleet efficiency: analyzing and validating before reaching for a model, and designing the experiments that tell you whether a change really worked. You lean on AI tools as a daily multiplier, but treat them as a fast, fallible collaborator, not an oracle: the bar on your judgment goes up, not down, and you're the one who catches the model when it's confidently wrong.

What we look for

• 2+ years across data science and data engineering: equally at home modeling a pipeline and reasoning about whether a metric is true
• Designs and owns streaming, event-driven pipelines that power live decisions: ETA, occupancy, routing, fraud
• Solves open-ended problems with classical ML, deep learning, or statistical inference, and can say why one over another
• Advanced SQL (Postgres), strong Python (pandas/numpy/sklearn), fluent in a BI tool; builds warehouse models other teams depend on
• Designs and evaluates complex experiments (A/B, quasi-experimental, switchback) and knows when a result is real
• Deep product and business instinct: analyzes and validates before reaching for a model, and directs AI well enough to override it when it is wrong

The assignment

Three mornings of raw GPS pings and trip/booking events from a handful of Mumbai routes. The question ops needs answered live: is route X running late right now, and by how much? Design the data model, build a runnable slice that ingests the pings and answers it, and surface what the data actually reveals, including the cases that look like lateness but aren't. A working slice beats a perfect diagram. Lean on your agent fully; we read the transcript too.

Up to a weekend; we expect 8–12 focused hours. Going well past that is a signal we'd rather not see; a tight, defensible slice beats a half-built platform. If you run out of road, stop and use the memo to tell us what you'd do next.