Smartrouter Project
Documentation
One landing page for all the Smartrouter modeling docs — start with the high-level overview, then drill into each training pipeline's README in the DATA-Athena-Snowflake repo. Engineering notes, design decisions, and feature definitions live alongside the training code.
📘Overview
High-level modeling overview — the canonical narrative for what we model, why, and how the pieces fit together.
Google Doc
Modeling Overview — Smartrouter
Top-level narrative covering the modeling stack: data pipelines, training architecture, model registry, deployment story, and how the three production training pipelines (Volaris DNN, Decline Optimizer, MCO) relate to each other. Start here for any new contributor.
docs.google.com/document/d/17oqSuWxmrqRjrmpjycmU1IKbOh8yJ61KV0Io2jhRo5Q
Open Google Doc →
🎥Tech Talk
Recorded walkthrough of the Volaris DNN training and serving architecture — start here if you want the spoken narrative alongside the code.
Video not loading? Open in Google Drive ↗
⚙️Training Pipelines
Three production training pipelines, each living in
DATA-Athena-Snowflake/src/training/<name>. Each README covers data sources, feature engineering, model architecture, evaluation, and the promotion path.Training Pipeline · GitHub
🤖 Volaris DNN — Processor Selector
TensorFlow per-processor independent classifiers that pick which PSP to route each Volaris transaction to. Currently in production (model_type
processor_selector, version dnn:v60:tf_per_processor). Heads for MIT_BULK, WORLDPAY, ELAVON, AMEX.DUNA-E-Commmerce / DATA-Athena-Snowflake / qa / src/training/volaris
Open training folder →
Training Pipeline · README
⚠️ Decline Optimizer
Analyzes decline patterns and surfaces actionable rules — which BIN/processor/error-code combos are reliably losing, and what to do about them (retry strategy, processor swap, hard-block).
DUNA-E-Commmerce / DATA-Athena-Snowflake / qa / src/training/decline_analyzer/README.md
Open README →
Training Pipeline · README
✉️ MCO — Message Config Optimizer
Predicts which authorization-message field tweaks (CIT/MIT toggle, AVS, MCC, soft-descriptor, etc.) maximize approval probability for each transaction. Multi-head architecture, one head per patchable field.
DUNA-E-Commmerce / DATA-Athena-Snowflake / qa / src/training/message_config_optimizer/README.md
Open README →