AWS re:Invent 2025 - Enterprise AI: Now, Next, Future (ANT202)

Enterprise AI: Now, Next, Future

Automotive Manufacturing Challenges

• Traditional automotive manufacturers face long development cycles of 4-5 years, while newer Chinese manufacturers have reduced this to 2 years.
• The R&D process is highly complex, involving engineering, efficiency, safety, and requiring significant expertise from senior automotive engineers.
• The goal is to use AI to speed up the R&D process and enable junior engineers to be as productive as senior experts.

Leveraging Unstructured Data and Time Series Analytics

• The starting point is the vast amount of specification documents containing tolerances, benchmarks, and performance requirements.
• These documents are converted into vector embeddings using Nvidia's Neotron services, enabling semantic search and retrieval based on meaning rather than keyword matching.
• Sensor data from the sophisticated test track, generating terabytes of time series data, is processed directly in the Teradata database using signal processing techniques.
• By combining the unstructured document data and structured sensor data, the system can answer complex queries that draw insights from both data sources.

Ensuring Data Privacy and Security

• The sensitive R&D data cannot be exposed to competitors, so the solution leverages AWS Bedrock and the "private GenAI" concept to keep all processing within the company's network.
• This allows using large language models like Claude for advanced querying and analysis without risking data leakage.

Transitioning to Automated Agents

• The next step is to move from augmented intelligence to automated agents that can perform tasks with minimal human intervention.
• This requires building out an ecosystem of components, including the agent framework, large language model, guardrails models, communication protocols, knowledge platform, and integration with other necessary tools.
• The agents need access to the data sources, analytics capabilities, and the ability to bring their own models (BYOM) for specialized tasks.
• To address the challenge of "tool overload" where agents get confused by too many options, the solution proposes using task-specific MCP servers to restrict the available tools.

Key Takeaways and Future Outlook

• Unstructured data is an untapped gold mine that can be unlocked through vector stores and advanced natural language processing.
• Hybrid environments combining on-premises and cloud infrastructure are common, and the solution needs to work seamlessly across these environments.
• The focus should be on building expert agents that can autonomously perform tasks and suggest solutions, potentially with humans making the final decisions.

Technical Details

• Neotron services from Nvidia are used for vector embedding and retrieval of unstructured data.
• The Neotron-powered extraction engine can parse PDFs at a rate of 30 pages per second per GPU, enabling the processing of 1.3 million pages per week per GPU.
• The Neotron models and OCR provide 50% more accurate results compared to open-source alternatives.
• The solution leverages AWS Bedrock and the "private GenAI" concept to ensure data privacy and security.
• The agent framework, communication protocols, and knowledge platform are built on top of the Teradata enterprise data platform.

Business Impact and Use Cases

• The AI-powered R&D process enables junior engineers to be as productive as senior experts, significantly improving R&D efficiency.
• The ability to quickly analyze vast amounts of unstructured data and sensor telemetry allows for faster identification of performance issues and optimization opportunities.
• The automated agent-based system can be applied across various industries, not just automotive manufacturing, to streamline complex, data-intensive processes and unlock new insights.