Price: 0
Number of applications: 11
11.09.26 (inclusive)
Piecework
Idea
ICT tasks
Общественное питание, доставка готовой еды (дарк-китчены)
Intelligent control systems
Software/ IS
The dark kitchen network operates according to the "central workshop → kitchens" scheme: the workshop produces semi-finished products (PF) with a shelf life of 24-72 hours and ships them on flights 1-2 times a day, the kitchen collects dishes from them to order. Today, the volume of production and shipments is determined by the chef's request "for last week". The error goes both ways: revaluation → PF is not consumed until the end of the shelf life → write-offs in the workshop and in the kitchen; underestimation → the dish goes to the stop list in the peak slot → missed orders and a drop in repeat orders. At the same time, the demand for this dish ceases to be observed (truncated), the following forecast is underestimated even more — the error self-intensifies. The task is not limited to forecasting sales. We need to move from forecasting the demand for dishes to a shift assignment for the workshop and an invoice for each kitchen in PF units, taking into account expiration dates, leftovers and flight schedules.
reduction of write—offs of PF and finished products as a% of the cost of shipments - target -25...40% to the baseline; reduction of the share of meals in the stop list during peak hours - target -30...50%; reduction of missed orders and cancellations due to unavailability of the dish; reduction of PF residues in kitchens with the same level of service; the application is generated automatically — the boss confirms or edits, not compiles.
Kulumbaev Kairzhan Serikbayuly
Purpose and description of task (project)
The system should answer one question daily (plus intraday adjustments): how much of each semi-finished product should be produced in the workshop and how much should be shipped to each kitchen on each flight. Demand forecast is at the level of "kitchen × dish × slot 30-60 min", the horizon is up to 3 days. Factors: history, weekly and intraday seasonality, holidays and weather, prices, promos, marketing campaigns, cannibalization between dishes, menu changes. The restoration of suppressed demand is the correction of the history for periods of stop lists and unavailability. Without this, the model learns from the consequences of its own mistakes. Technical mapping is the translation of the forecast of dishes into the need for PF: the norms of the bookmark, the output after processing, the total PF included in several dishes. Calculation of a batch with a clear economic balance: the price of the deficit (lost margin + stop list effect on the user) versus the price of the surplus (cost x probability of not selling by the end of the term). The level of service is differentiated: higher for locomotive dishes, lower for the tail of the assortment. Restrictions: leftovers in the kitchen and on the way by batch (FEFO), shelf life, flight schedule and capacity, packing multiplicity, line capacity and production window of the workshop, kitchen refrigerating capacities. System output: shift assignment to the shop; invoice for the kitchen for each flight; list of PF and dishes at risk (limit sales / launch promo for the deductible); explanation of the figure ("+18 servings to the average: Friday + promo for combo").
Note
Acceptance criteria WAPE forecast of PF demand on the flight horizon — necessarily against a simple baseline (moving average of slots + buffer); ML is accepted only with significant superiority; bias is near zero — systematic upward skewing "just in case" is not counted; pilot 4-6 kitchens, 4-6 weeks, switchback or control group; both metrics at the same time — write-offs and stop lists; improvement of one at the expense of the other is not accepted; fill rate: the percentage of application items shipped in full by the workshop. ____ The recommendation system: the shop chief and the kitchen manager see the recommendation with an explanation and can change it. The discrepancies "recommendation → fact of production" are logged and go into training.
Whatsapp
