AI-Powered Supply Chain Visibility and ETA Prediction System

toyota_/_ibm · Automotive · 2025 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Automotive vehicle supply chain visibility

Toyota wanted more accurate and transparent vehicle ETA predictions across the order-to-delivery journey so dealers and customers could understand where a vehicle is, whether delivery is at risk, and what alternatives exist when delays occur.

Constraints:

• vehicle delivery spans disconnected ordering, scheduling, manufacturing, quality, transportation, inspection, and dealer-delivery stages
• legacy systems built 20-30 years ago, file-based batch processes, and siloed data slow visibility
• mainframe data extraction cannot overload live operational systems
• ETA predictions must combine real-time events, batch data, data-quality checks, and machine-learning inference
• dealers and end customers need different views without losing consistency

Architecture

Toyota, IBM, and AWS built a supply-chain visibility architecture that moves mainframe and legacy vehicle-event data into AWS through CDC and SFTP, streams it through Kafka, enriches it with Flink and rules services, and stores it in domain-oriented operational stores. SageMaker Wrangler cleans and engineers features; regression, time-series, and classification models such as XGBoost, AdaBoost, and Random Forest estimate ETA at each vehicle journey node. Batch transform precomputes inference every four hours for near-real-time use. Downstream dealer/customer applications expose a pizza-tracker style vehicle journey, while a Bedrock-based CUBE Assist/ETA chatbot lets users ask natural-language status questions.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  legacy-source-systems["legacy-source-systems"] --> cdc-ingestion["cdc-ingestion"]
  legacy-source-systems["legacy-source-systems"] --> sftp-batch-ingestion["sftp-batch-ingestion"]
  cdc-ingestion["cdc-ingestion"] --> aurora-landing-store["aurora-landing-store"]
  cdc-ingestion["cdc-ingestion"] --> kafka-event-backbone["kafka-event-backbone"]
  sftp-batch-ingestion["sftp-batch-ingestion"] --> kafka-event-backbone["kafka-event-backbone"]
  aurora-landing-store["aurora-landing-store"] --> kafka-event-backbone["kafka-event-backbone"]
  kafka-event-backbone["kafka-event-backbone"] --> flink-enrichment-layer["flink-enrichment-layer"]
  kafka-event-backbone["kafka-event-backbone"] --> domain-data-stores["domain-data-stores"]
  kafka-event-backbone["kafka-event-backbone"] --> sagemaker-ml-pipeline["sagemaker-ml-pipeline"]
  flink-enrichment-layer["flink-enrichment-layer"] --> domain-data-stores["domain-data-stores"]
  flink-enrichment-layer["flink-enrichment-layer"] --> exception-management["exception-management"]
  domain-data-stores["domain-data-stores"] --> sagemaker-ml-pipeline["sagemaker-ml-pipeline"]
  domain-data-stores["domain-data-stores"] --> dealer-customer-apps["dealer-customer-apps"]
  domain-data-stores["domain-data-stores"] --> cube-assist-agent["cube-assist-agent"]
  sagemaker-ml-pipeline["sagemaker-ml-pipeline"] --> dealer-customer-apps["dealer-customer-apps"]
  dealer-customer-apps["dealer-customer-apps"] --> cube-assist-agent["cube-assist-agent"]
  cube-assist-agent["cube-assist-agent"] --> agentcore-orchestrator["agentcore-orchestrator"]
  exception-management["exception-management"] --> dealer-customer-apps["dealer-customer-apps"]

Components

Data flow

Vehicle lifecycle events originate in mainframe, ERP, national vehicle, dealer, port, and logistics systems. CDC captures mainframe changes without live-querying the mainframe, while SFTP carries offline batch feeds. Both paths publish events into Kafka topics. Flink and separated rules engines enrich and transform the stream; data quality, lineage, retry paths, domain stores, REST APIs, and pub/sub expose events to downstream applications. ML pipelines land historical/event data in databases and S3, clean it with SageMaker Wrangler, train and compare candidate ETA models, precompute batch inference every four hours, and serve VIN-specific ETA and journey status to applications and chatbot agents.

Why this architecture

The architecture separates data capture, event streaming, transformation, domain stores, ML inference, customer applications, and agentic experiences because the vehicle journey crosses old systems, third-party logistics, and different customer needs. CDC protects mainframe performance; Kafka gives real-time event distribution; Flink and separate rules improve maintainability; SageMaker handles feature engineering and ETA model operations; batch inference every four hours provides near-real-time readiness; and the agent layer turns the resulting visibility into natural-language interactions while keeping humans in exception management.

Models used: Amazon SageMaker · SageMaker Wrangler · XGBoost · AdaBoost · Random Forest · regression models · time-series forecasting · classification models · Amazon Bedrock · Anthropic model · Amazon Nova · AgentCore · MCP server · A2A protocol

Scale

• The talk was a roughly 45-minute re:Invent session.
  “And our rough talk will be around 45 minute.”transcript, stated
• The number of vehicle configurations doubled in five years.
  “just in the last five years, the number of vehicle configurations has doubled”transcript, stated
• Top six automotive manufacturers reduced 2025 profit estimates by over $25 billion due to tariffs.
  “Just in 2025, the top six automotive manufacturers had to reduce their profit estimate by over 25 billion because of tariffs.”transcript, stated
• 94% of companies have been affected by supply chain disruptions.
  “94% of all companies have been affected by supply chain disruptions”transcript, stated
• The average new vehicle transaction price in the US crossed $50,000 in September.
  “the average transaction price of a new vehicle in the US was over $50,000”transcript, stated
• Legacy supply-chain systems were described as 20-30 years old.
  “Some of these systems were built probably 20, 30 years ago.”transcript, stated
• Toyota cited about 1,200 unique transit links.
  “We have unique transit links, right? We have some numbers there, about 1,200 there.”transcript, stated
• The model precomputes inference every four hours for near-real-time ETA readiness.
  “if you actually do the inferencing every four hours, you are almost ready for the next event to come in.”transcript, stated
• Vehicle journey examples include one-day build windows, 0.5-day quality, two-day yard/shipping readiness, and multi-hour or day-long transportation moves.
  “that can take from one one day to maybe a couple of days, right? Here are the quality. We say, hey, this should take X amount of time.”transcript, stated

  “can take up to two days.”transcript, stated

Alternatives considered (roads not taken)

Live querying the mainframe for operational data

Rejected because: It could adversely affect mainframe performance, so the architecture uses CDC instead.

Batch/file-based end-to-end updates

Rejected because: Manual and batch processes create hours or days of delay before customers learn about logistics issues.

Static one-size-fits-all vehicle status view

Rejected because: Different dealer and customer users want either detailed milestones or only the final arrival answer.

Single preselected model/runtime for the agent layer

Rejected because: AgentCore is attractive because it supports different models, frameworks, protocols, runtime, observability, and identity.

Patterns

batch-offline-processing (ml operations)

The ETA system precomputes batch inference every four hours so downstream applications can respond near-real-time when new vehicle events arrive.

observability-tracing (data operations)

The architecture explicitly includes lineage, root-cause traceability, model monitor drift checks, exception retry paths, and alerts for the right team.

multi-agent-orchestration (agent architecture)

The next agentic layer frames ETA, port processing, and vehicle-system checks as agents coordinated by an orchestrator through AgentCore/A2A/MCP-style infrastructure.

human-in-the-loop-gate (workflow)

The operating model keeps humans focused on exceptions while automation handles the routine journey and ETA processing.

Tradeoffs

Use CDC for mainframe replication.

Chose Change-log capture and replication over Live querying the mainframe.

Rationale: CDC protects mainframe performance while allowing event data to enter cloud streams.

Precompute ETA inference periodically.

Chose Four-hour batch transform readiness over Only computing inference from scratch on every event.

Rationale: The system can serve near-real-time updates by reusing precomputed scores when new events match.

Represent ETA as a window rather than one brittle date.

Chose Start/end duration window with confidence and status over Single exact arrival date.

Rationale: Toyota wants a tight window and high accuracy, but acknowledges uncertainty changes as a vehicle moves through the pipeline.

Provide both detailed milestones and simple arrival answers.

Chose Personalized journey maps and multiple views over One universal customer-facing status view.

Rationale: Some users want detailed milestone progression while others only care when the vehicle arrives.

Gotchas (what bit them)

Legacy and siloed systems delay customer visibility.

Root cause: Different stages run on independent systems, inconsistent data models, manual processes, and old platforms that cannot connect easily to cloud or AI workflows.

Resolution: Move event data into a unified, event-driven AWS architecture with CDC, Kafka, polyglot storage, APIs, and downstream workflow enablement.

Mainframe extraction can harm source-system performance.

Root cause: Live queries against operational mainframes are risky for existing transaction workloads.

Resolution: Use CDC sidecar journal/change-log capture and offline SFTP batch paths rather than live mainframe queries.

ETA quality depends on data intimacy and variability control.

Root cause: Variable, low-quality, or poorly understood logistics data makes machine-learning predictions less reliable.

Resolution: Invest in data quality, feature engineering, model monitoring, and understanding the data before relying on ETA predictions.

Evaluation

Toyota evaluates ETA prediction through accuracy, confidence scores, ETA window size, schedule status, historical performance, model drift, and whether dealer/customer views improve visibility and satisfaction.

Metrics: ETA accuracy · confidence scores · duration-window size · on schedule / at risk / delayed status · model drift · customer satisfaction · replacement vehicle opportunities

Outcomes

• Toyota built a vehicle pipeline experience that gives dealers and customers ETA status, lifecycle milestones, documents, ETA history, and vehicle journey context.
• The ML pipeline calculates node-level and dealership-level ETAs and refreshes inference every four hours for near-real-time use.
• The team is moving toward an agentic CUBE Assist experience for instant supply-chain visibility, proactive delay detection, replacement options, and human-in-the-loop decisions.

The case cites industry-wide disruption affecting 94% of companies, Toyota journey complexity across seven stages and about 1,200 transit links, four-hour inference refreshes, and vehicle prices over $50,000.