RetailCase Study

Swiggy: MIMO Deep Learning Cuts Delivery Prediction Training Time 80%

Swiggy: MIMO Deep Learning Cuts Delivery Prediction Training Time 80%

·11 min read
SwiggyFood Delivery & Quick Commerce5M daily orders · 500K restaurants · 1M delivery partners · 50M monthly active users

Key People & Companies

MIMO Deep Learning Model

Swiggy · Internal

Swiggy

Entity Embeddings

Swiggy · Framework

ETA Model

Swiggy · Internal

Soumyajyoti Banerjee

Data Scientist at Swiggy

TensorFlow

Google · Framework

DSP (Data Science Platform)

Swiggy · Platform

Sunil Rathee

Swiggy

+ 8 more entities in the full study

Key Results

  • Training time reduced to one-fifth of previous approaches, with training memory footprint cut by approximately 50%. This allowed models to train in-memory rather than requiring disk-based operations, according to Swiggy's engineering blog.
  • Order-to-assignment (O2A) prediction accuracy improved by nearly 30%, helping the assignment system allocate delivery executives faster and more efficiently. As the team noted, 'This helped our exception management systems to become more efficient at recovering from delays in DE assignment.'
  • Overall delivery time prediction (O2R) mean absolute error decreased by 5% after introducing in-model embeddings, improving customer experience through more accurate delivery promises.
  • + 4 more results inside

Having separate networks for each of the outputs was not only inefficient to train and deploy but also left unrealised performance benefits on the table.

Shubham Grover, Data Scientist at Swiggy

Sign in to unlock the full case study

Get the complete analysis, all results, expert quotes, and entity profiles — free.