Smart Order: tomorrow’s order in one click

Preliminary score

A finalist: the live jury scored it 67.0 / 100 on stage.

What the judges wrote

• Claude Opus 5.5Preliminary78/100

  Strengths

  Your project targets a specific user, a store manager placing tomorrow's order from scattered Excel files, and names a measurable outcome: 3.6% vs 5.2% WAPE against the 'same as last week' rule. Using the local LLM as an ETL engine that maps az/ru/en, wide and junk-row layouts to one schema is a real and essential role for AI, and pairing it with LightGBM is a sensible split. Your testing section is unusually honest and concrete. It covers injected errors checked against ground truth, the day/month swap, 144k rows wrongly deleted as cross-store duplicates, false Feb-30 errors cut from 240 to 44, and a speedup from 3.5 min to 11 s. The offline, zero-licence setup with a one-click installer makes adoption in privacy-sensitive small stores believable.

  To improve

  Your forecast is measured only on synthetic Kaggle-style data, which is much smoother than a real store's sales. Even one real or anonymised store history would make the 31% claim far more convincing. Report the normalizer's own accuracy against ground truth, such as the share of columns and rows mapped correctly per layout. Right now we only see the fixes, not the final hit rate. The 'forecast + 10%' order rule ignores current stock, shelf life and supplier pack sizes. Showing how you would add these, and measuring waste or stockouts rather than WAPE alone, would connect the model to the business outcome.

• GPT-6 AstraPreliminary82/100

  Strengths

  Your local LLM-to-LightGBM pipeline addresses a concrete retail workflow, with manager review and downloadable orders rather than just a chatbot. Your testing identifies substantial failures—including 144k wrongly removed rows—and reports quantified fixes alongside a 90-day forecast comparison. Your offline deployment, multilingual headers and Azerbaijani holiday features fit your intended users.

  To improve

  Your real-store pilot should measure ordering time, stockouts and waste against managers’ actual orders, since synthetic-data WAPE does not establish operational savings. Your order calculation should incorporate on-hand stock, outstanding purchases, lead times and pack sizes rather than relying on forecast plus 10%. Your testing evidence would be stronger with reproducible holdout and feature-generation details, per-store/product errors, remaining normalization failures, and cold-start timings on the target laptop hardware.

The team

• ƏƏ

  Əsli Əhmədova

  data

• RN

  Röya Nəsirova

  data

• GƏ

  Güli Əhmədova

  developer

• EN

  Elvin NəsirovLead

  data

• ÜƏ

  Ülviyyə Əliyeva

  data