Reranked a tutoring marketplace with machine learning.
Led Search and Discovery; reranked the marketplace with AI/ML and moved the two numbers that matter.
01 · The problem
What Preply actually needed.
Preply is a global language-tutoring marketplace. Learners search for a tutor, and the order tutors appear in is the single biggest lever on whether a learner books and whether the company makes money. When I took over Search and Discovery, that order ran on the platform's initial algorithm, not on a model trained to predict what actually leads to a booking. On a marketplace, ordering is the product: every position allocates supply to demand, and a ranking tuned for the wrong signals quietly leaves conversion and GMV on the table at scale.
02 · Context and insight
The reframe that set the direction.
Marketplace ranking is an information-retrieval problem before it's a UI problem. The initial algorithm encoded reasonable heuristics (price, rating, response time) but never learned the actual weighting from behavior. I reframed ranking as a prediction problem: order tutors by their predicted likelihood of a successful booking, learned from real behavior. That carried a specific risk worth naming. Conversion and GMV can pull in different directions: optimizing purely for conversion pushes toward cheaper lessons, while optimizing GMV alone can hurt match quality. Moving both up from one core change was the result that mattered.
03, The approach
The decisions that mattered.
Treat ranking as the marketplace's core allocation decision
I anchored the team on one question: of all the tutors who match a search, in what order should they appear to maximize successful bookings? Framing ranking as supply-to-demand allocation, not a list to be styled, set the priorities. The initial algorithm was the baseline to beat, and the bar for replacing it was a measured lift in conversion, not a subjective sense that results 'looked more relevant.'
Pick the objective deliberately, knowing conversion and GMV can diverge
The senior tradeoff was the optimization target. A ranking tuned only for conversion favors cheaper lessons. One tuned only for GMV surfaces tutors learners won't book. I chose predicted successful bookings as the primary objective with GMV as a guardrail, on the thesis that better matches lift both. The outcome validated it: conversion rose 7% and GMV 4% from the same change. The deliberate part was accepting those numbers could have split and committing anyway.
Prove it with experiments before trusting it marketplace-wide
Because ranking touches every search, I treated rollout as risk management. We validated the learned ranking against the incumbent through controlled experiments, so the 7% conversion and 4% GMV lifts were measured against a live baseline, not modeled in a notebook. That discipline made it safe to apply the change across the whole marketplace rather than a slice, while keeping supply health in view so the win wasn't borrowed from the long-term health of the tutor base.
04 · How it's built
Close to the stack, not above it.
This was a product-leadership role, not an IC engineering one. The Search and Discovery ML and engineering team did the building while I owned the problem framing, the objective, the labels, and the rollout bar. The hard technical calls were product calls in disguise: what counts as a 'successful booking' for training, which signals the model is allowed to weigh, and how far it can deviate from the trusted initial algorithm before a human reviews the result.
Replacing the initial ranking algorithm lifted conversion 7% across the entire marketplace and GMV 4%. Both numbers came from one core change applied platform-wide, and that is the point: because ranking sits in front of every search, a single well-targeted improvement compounds across all demand. Moving conversion and GMV together, when those objectives can easily pull apart, is the result I'm proudest of from this role.
What I’d carry forward
On a marketplace, choosing the optimization target is the real decision. The model is downstream of it. Conversion and GMV happened to move together here, but I went in knowing they might not, and that framing kept the team honest about what we were trading. The lasting risk in any ranking win is starving long-term supply health for short-term bookings: the axis I'd watch hardest over a longer horizon.