Scaling Intelligence at the Physical Edge: Lessons from Amazon’s Logistics Automation

In this conversation, Keerti Kulkarni explains the shift from treating physical logistics data as noise to treating it as a primary operational signal. By integrating AI agents into the physical dock door environment, her team at Amazon moved beyond basic dashboards to proactive, conversational troubleshooting. The takeaway is that the most leverage in large systems comes from creating tight feedback loops between physical operators and the data layer, rather than just collecting more information. For technical leaders, this conversation shows that scaling operations requires a foundation of rigorous data reconciliation, which then allows for the use of high-leverage AI tools. Teams that prioritize data pipeline integrity over AI novelty will gain a significant advantage in managing complex, real world operations.

The Hidden Cost of Fast Data

Most engineering teams treat telemetry as a byproduct to be smoothed over. Kulkarni argues this is a mistake. When you treat physical events as noise, you lose the ability to perform accurate root cause analysis on safety or hardware failures. By tagging events with both site clock and ingest clock timestamps, her team performs explicit reconciliation to ensure data integrity before it reaches the analysis layer.

> We treat the physical layer as a first class telemetry, not as a noise which is to be smooth later. So each vehicle entry point reports state and we tag events with side clock and ingest clock. So if there is any difference between this then there are major effects seen downstream.
>
> -- Keerti Kulkarni

This approach creates immediate work, as designing a ledger with event keys and explicit summarization is labor intensive. However, it eliminates the debugging problems that occur when teams try to reconstruct state from inconsistent logs. Over time, this creates a reliable foundation that allows the system to scale without compounding technical debt.

Why the Obvious Fix Makes Things Worse

Conventional wisdom suggests that adding LLMs to a high volume system will solve visibility problems. Kulkarni notes that without strict vocabulary controls and structural constraints, the AI becomes a source of hallucination rather than clarity. Her team avoids this by attaching ledger event IDs to every AI generated summary. This forces the model to ground its output in actual, verifiable data.

> We are trying to basically limit the summary length. We are trying to as I said attach the ledger event and event IDs and the vocabulary. We are having very strict vocabulary control on the agentic framework as well.
>
> -- Keerti Kulkarni

This creates a human in the loop advantage. By providing operations teams with a structured, grounded summary, they can verify the root cause before acting. This prevents the system from bypassing the operators, ensuring that AI acts as a tool for human decision making rather than an opaque, automated box that creates more confusion than it resolves.

The 18 Month Payoff of a Boring Foundation

Kulkarni’s transition from static dashboards to an agentic framework was only possible because they first separated their environment from shared clusters and enforced strict cost attribution. By tagging Firehose streams by site, they turned cost into a signal for operational health. This enables a delayed payoff: they can now identify power users and sites with high failure rates, allowing for proactive, rather than reactive, training interventions.

This is where the system demonstrates its durability. By building a modular agentic framework, the team is moving toward a future where authentication and authorization logic can be shared across different logistics teams. This requires the patience to build for modularity now, even when a quick, monolithic fix would satisfy the immediate demand for a new feature.

Key Action Items

• Implement Dual Clock Tagging: Immediately begin tagging telemetry with both site clock and ingest clock timestamps to identify latency and ordering issues before they hit your analysis layer. (Immediate)
• Establish a First Class Telemetry Policy: Transition from treating logs as exhaust to treating them as a ledger. Prioritize reconciliation steps at every pipeline stage. (Over the next quarter)
• Enforce Grounded AI Outputs: If deploying LLMs for troubleshooting, mandate that every summary must include the raw event IDs or ledger references used to generate the conclusion. (Immediate)
• Granular Cost Attribution: Tag data streams by site or business unit to enable precise cost per assignment tracking. This creates the visibility needed to identify anomaly driven costs versus growth driven costs. (Next 3-6 months)
• Build Modular Agentic Frameworks: Instead of building monolithic AI agents, focus on modularizing authentication and role based access control so they can be reused across different logistics workflows. (12-18 months)
• Proactive Training Feedback Loops: Use your telemetry to identify patterns in operator failure and feed these insights directly into training programs rather than relying on generic site wide updates. (6-12 months)