AI-ifying operations in the Yandex Eda editorial team
Как мы ИИ-зируем операционку в редакции Яндекс Еды
B · Useful, limited A practitioner talk by the editor-in-chief of Yandex Eda with a clear stage-by-stage pipeline design and an evaluation method; the time savings are self-reported, partly from an unnamed stream, with no error rates or costs.
Summary based on the full transcript·Original: about 33 minutes to watch
In short
A 33-minute Russian conference talk (Stachka 2026, uploaded 28 October 2026) by Kira Kalimulina, editor-in-chief of Yandex Eda, the Russian food-delivery app, who joined in February 2026. Her team writes nearly all app texts and handles about 900 tickets a month. Editors and the CRM team first built, without developers, a "briefer" that chases managers for missing brief details and an AI editor for push notifications driven by a 40-page prompt. That version became the business case. She then designed a five-stage pipeline of skills (understand, plan, produce, check, retain knowledge) that one developer is now building; "more than half" is in place.
Key numbers
- ~900tracker tickets a month handled by the editorial team
- 2-3 hours to 15-30 minutestime to prepare push notifications before and after the first AI version (claimed 7-10x)
- 1,000tickets used to test the first version, each rule scored 0 or 1
- 200 h to 68 htime for a stream of 135 tasks before and after the pipeline
As reported by the source.
Russian·Video·opens youtube.com
What it reports
- First version: task information to AI-written brief to push texts to a human editor who approves or edits. The prompt ran to 40 pages with 12 large sections on punctuation, quotes and style; attempts to shorten it failed.
- It was tested on 1,000 tickets, each rule (for example, no bureaucratic language) scored 0 or 1. Early editor feedback was poor.
- Push preparation fell from 2-3 hours to 15-30 minutes; she claims a 7-10x cut. Freed time went into creative pushes, one of which was widely copied by other brands.
- Pipeline: AI handles task understanding and ticket updates (formerly a project manager's job); a planner assigns work by format, leave, load and story points. In production, a briefer queries managers, an editor skill writes, a checker tests for known recurring model errors, a human gives free-text feedback ("too dry"), a feedback skill records it, an improver rewrites, the checker also checks the human edit, and the final text goes into a base of approved texts for similar cases.
- Of the B2C tickets, almost 300 a month run through AI; the briefer sent about 600 requests for missing information (roughly two per ticket) that people used to send.
- On a separate stream she would not name, 135 tasks took 200 hours before and 68 hours after, "at least three times" faster.
- Freed capacity created new UX, SEO (four magazines; up to 3.5x views on some articles) and "knowledge" desks; she says headcount reduction is not a goal.
- The briefer first ran on a button nobody pressed; it now runs automatically once a day, starting 5-7 days before a task is due depending on format.
- Self-check error rates are not yet measured; business-case economics were withheld under NDA. On the model, she said the architecture works with whatever model a team uses.
How they got it
A conference talk with slides and audience Q&A; the speaker reports her own team's figures. The 1,000-ticket evaluation method is described, but not its pass rates. The digest is based on cleaned Russian auto-captions; the speaker's surname comes from the captions and is unverified.
Limits
- Self-reported; no baselines beyond the quoted hours, and the 135-task stream is unnamed.
- The pipeline is only partly built, so the figures mix the first version and partial implementation.
- No checker error rates, model name or costs.
- The library summary credits non-engineers with the five-stage pipeline; in the talk, editors built only the first version, and one developer is building the pipeline.
Why it matters for companies
The design is reusable outside editorial work: split the job into stages, give each a narrow skill, check model output against a list of known errors, capture the human reviewer's correction as structured feedback, and store approved outputs as cases. The briefer story is a common adoption lesson: a step that waits for someone to press a button does not run, so triggers should be automatic and scheduled to deadlines. The team built a working first version without developers and used it to win one developer for the full build.
Key passages in translation
Russian original
Сначала мы занимались этим всем 2-3 часа, потом мы скооперировались и оказалось, что всё это можно сделать там буквально за 15-30 минут.
At first all of this took us 2-3 hours; then we teamed up and it turned out it could all be done in literally 15-30 minutes.
Russian original
А там у нас, допустим, 135 задач, в среднем это занимает 200 часов, но после реализации а это занимает, ну, сильно меньше, 68 часов.
There, say, we have 135 tasks; on average that takes 200 hours, but after implementation it takes much less, 68 hours.
Russian original
изначально мы его делали по кнопке. [...] И, ну как бы не понятно, что никто не не не ну не включал её.
Originally we made it run on a button. [...] And, well, it turned out nobody ever switched it on.
Summary based on the full transcript. Read the original for the full argument and data.Watch the original on youtube.com