Online retention logic had been copied onto an offline model with lesson deposits. Active base, debt and critical debt weren't separated. A parent who stops coming rarely bothers to cancel formally, so the dashboard looked fine while classrooms emptied.

Refused to pick the prettier number. Took students absent for a whole month and traced what they did next — returned, topped up, sat in the system until lessons ran out. Wrote the definitions for analysts (active / debtor / critical, when "churn" is an event), got a weekly debtor report by office, cleaned the base. Then attached actions: a deposit top-up programme with a clear explanation of what the deposit is for, and Bandung's practice of billing everyone by the 10th rolled out to every branch.

A debtor registry and weekly follow-up that didn't exist before. In my self-review I reported retention of 96.9% in May 2026 versus roughly 89% earlier.

Teacher payouts ran ~40% above plan month after month. The accepted explanation was inaccurate planning, which nobody can manage.

Pulled monthly payout documents with the analysts, clarified how teachers are paid, and compared regular versus catch-up lessons. In several branches catch-ups exceeded regular lessons. Turned it into a rule: catch-up cost capped at 20% of regular lesson cost in June, 10% from July — written into lead tutors' goals, with learning quality as the constraint on how far to cut.

One branch manager reported catch-ups down from 140 to 8 lessons a month.

Lead volume grew, sales didn't. Lead-to-booking fell, and the story was "offline sources are lower quality".

Split the traffic-mix effect from the processing effect: compared conversion inside each source, by city, by stage. Online was degrading as well. The mechanism: delayed first contact, a mostly new booking team, and an SLA/QC scheme that rewarded working old leads. Reworked priorities toward speed, scripts and slot allocation, and set a routine: every lead that didn't reach booking gets a reason, reasons get grouped, each group gets a small fix. Separately, removed a post-booking budget question that was suppressing show-ups (28 May) and watched booking-to-attendance afterwards.

Diagnosis and the operating routine. Attendance-to-sale for 2026 YTD was compared at 31.7% against a 2025 average of 27.1% in July discussions.

Functions and processes were tangled between two parallel managers; people's strengths were used unevenly; the weekly meeting was a reading of numbers with answers arriving a week late.

Had both managers describe their processes and what they wanted to own, mapped the work by function, and proposed the structure: branch directors run operations and the school's P&L; functional owners run sales, standards, projects and support. Later added a Student Experience & Retention Lead. Changed the rhythm rather than the slide template: Friday we open the dashboard together with no prepared slides, split the questions between marketing and booking, and by Wednesday the owners come back with a plan. When a manager left, his processes went into three piles — to branches, to the sales lead, and "nobody yet, document it thoroughly".

By November the matrix was the working principle, not an idea. The Friday meeting was in calendars from 15 October. Onboarding got a team-structure overview after an engagement survey showed new hires didn't understand how the functions connected — a gap I owned rather than blamed on "low engagement".

An ambitious model and the actual numbers had drifted apart. Opening more cities on unverified assumptions would scale the loss.

Rebuilt the forecast with finance: correct historical conversions, the real lag to second payment, branch result shown separately from CAPEX and HQ allocations. Framed the question for the board as "under what conditions does the model work, and which of them have we already confirmed with facts" — and proposed letting the changes already made show their effect before expanding. Set up research on Makassar, Batam, Bogor, Pekanbaru, Palembang, Balikpapan, Samarinda, Bekasi and Padang: paying families, schools and residential density, competitors, site availability, branch economics, a strong local lead — plus our own online-lead history and the behaviour of online students from those cities. For Madiun, checked that free rent isn't the same as a viable branch. For Surabaya West: small active base, renewing the lease meant roughly $4–4.5k of loss, so we moved the students and closed it — splitting the legal closure from a late water bill so one didn't hold up the other.

Medan launched as the seventh city; the local branch manager gets the credit for the site and the opening, I ran the goals, the commercial follow-through and the launch communication. West Branch was being vacated by July.

Kazakhstan's share of bad orders (cancellations, wrong orders, bad food) sat at 8.9% against a target of 7.5%. Nine months of partner work — auto-blocks, integrations, stop-lists on cancellation, photo control — had taken restaurants about as far as they could go.

Decomposed the rate by partner and cause, and built a quarter plan with a number on each initiative: long-tail contact and schedule refresh, auto-stop for "dish missing", KFC integration, promo-code-only compensation for KFC, fraud excluded from the wrong-order calculation, "dead souls" auto-disabling of inactive places. Courier fraud on "failed call" cancellations — couriers phoning support, "can't reach the restaurant", cancel — was removed by taking the call out of the flow: the courier hands the phone to a restaurant employee. On the client side, 76% of accounts sat in an unrated "grey" zone; set a target of ≤20% and added a loss-making-client rule: LTV minus cash, promo and logistics compensation below zero means no more cash refunds. Hired a content collector for small cities. Started removing the 200 lowest-performing partners.

Failed-call restaurant cancellations 447 → 336. Menu completeness in the smallest city 60% → 72% in two weeks. On my CV for the period: partner cancellations down to 0.73%, bad-food refunds cut by 25%, wrong orders down 1 pp, bad-food incidents down 17 pp, courier pick-up wait down 1 minute, service opened in 8 new cities. The 40-person team was named innovation leader across international Yandex Eats.

Too much work in progress, constant priority conflicts, no time ever left for legacy, and opaque progress that had customers blaming engineering and engineering blaming customers.

One board for all development work; nothing gets worked on that isn't on it. WIP limits: backlog 40 (bugs 10 / features 20 / legacy 10), to-do 10, one task per developer. A single product owner sets priority. Every issue carries two lines from its author — size of the possible economic effect and the probability it happens — because a $10 issue shouldn't cost everyone's time. Launched 21 December 2022 after two weeks of collecting objections and cleaning the board.

A 5-engineer team with a visible queue and a predictable "when": on any priority task, an estimate within hours. I also owned the Stories widget as a product and the trigger-email app for Shopify.

A dark-kitchen startup (Varlamov.Est, Carrots.Kitchen, Vedomosti.Lunch and others) needed to open fast and cheap, then run a hundred-plus cooks across sites without the wheels coming off.

Opened 15 kitchens end to end — search, rental agreements, build and equipment — at the lowest opening cost I'm aware of in the segment, under $50k (3 million roubles) per kitchen. Opened sales channels (catery.ru, Yandex Lavka, WeWork) and built our own brand, Carrots: agency, site, apps, fulfilment, promotion; the brand was nominated for G8 Awards. Moved from Director of Business Development to COO and ran the team of 100+.

Order hand-out time 22 → 15 minutes; output per cook 16 → 21 orders a day — both beyond the planned targets.