Branch managers became responsible for their school’s operations and P&L. Functional leads had the authority to make decisions in their area and were responsible for testing new approaches, identifying what worked and putting those practices into use across all branches.

Among the changes we introduced:

• Shared sales script and training: A2S increased by 5.2 percentage points
• Shift towards larger packages: increased AOV by $85
• Merged underfilled groups to free up peak-time slots for new students without adding classrooms

YoY: January–August 2026 vs. the same period in 2025.

With the analytics team, we compared expected teaching costs based on scheduled lessons and teacher rates with actual payments. Branch-level analysis showed that much of the additional cost came from catch-up classes - free make-up lessons for students who had missed regular classes. In some branches, there were almost twice as many catch-up lessons as regular lessons.

We redesigned the catch-up model: free sessions were limited to valid exceptions, handled by lead tutors within their existing workload, and grouped whenever students needed to cover the same material.

This substantially reduced catch-up volume and brought teacher spending back under control.

I led a review of how we measured retention and shifted the model towards actual payments. Students with zero or negative lesson balances were no longer counted as retained, revealing how much of the problem had previously been hidden.

A new branch-level dashboard showed which students needed payment follow-up, with direct links to their records. Teams could identify payment issues early and act on them.

The share of students behind on payments fell from 52% to 13%, while retention improved by 29 percentage points YoY.

I compared performance by month, quarter and year, and looked at how branches behaved at different stages of maturity. The analysis showed that first-year growth could not simply be carried forward. Each school served a relatively small local area: new branches could grow quickly as they captured this audience, but over time growth naturally slowed as the pool of potential new students became more limited.

This gave us a clear basis to reforecast after Q1. Together with finance, we rebuilt the model using historical conversion rates, actual payment timing and updated costs. We also accounted for how changes in package mix affected revenue timing: larger packages increased upfront revenue from primary payments but pushed secondary payments further out.

The revised model separated what the existing network could realistically deliver from growth that would need new investment. On that basis we held off on further expansion until the changes already underway had time to show results, transferred students out of the loss-making Surabaya West branch and closed the site, and opened Medan as our seventh city.

My personal favourite was “Dead Souls”: finding restaurants that had closed but were still listed on the platform. Customers could place orders that would never be fulfilled. Removing these restaurants helped prevent failed orders and contributed to a record low in restaurant cancellations in Kazakhstan.

The biggest financial impact, though, came from changing how support handled food-quality feedback. I read hundreds of support tickets and proposed a change to the response script to separate customer reviews from refund requests. This saved approximately RUB 0.20 per order across Kazakhstan.

After two weeks of discussing objections and clearing the backlog, I introduced one board for the whole development team. Each developer worked on one task at a time. The backlog had limits and reserved space for maintenance, and one product owner set priorities.

Anyone requesting work had to estimate its potential financial value and the likelihood of that value being realised. The five developers had a visible queue, and priority requests could receive an estimate within hours. Alongside this, I managed the Stories widget and a Shopify app for automated email campaigns.

I went on to run operations with more than 100 cooks. During that period, the time to get an order out fell from 22 to 15 minutes, and daily output per cook rose from 16 to 21 orders.

I also built the Carrots brand with an agency, covering its website, apps, fulfilment and promotion, and opened sales channels through Catery, Yandex Lavka and WeWork.