In last-mile delivery, the critical operations occur within minutes of package handoff. Modern vehicles, continuously connected, telemetrically rich, and policy-aware, function as mobile, verifiable endpoints. Identity, context, and sensor data co-locate with that endpoint, ensuring access is granted only to authenticated entities in the correct spatial and temporal context. Within these defined guardrails, machine learning models can execute decisions at the curb with speed, auditability, and deterministic recovery paths.
Bhavna Hirani, Software Development Manager at Autodesk, applies this rigor from design through deployment. As Associate Editor for the Sarcouncil Journal of Innovative Science, she views reproducibility and peer verification as engineering imperatives. Her principle is constant: bind access to identity, measure model outputs against operational data, and promote features only after they sustain performance in production environments. This same mindset guided her work shipping Amazon Key In-Car Delivery, which integrated scoped credentials and cryptographically signed access logs—principles she continues to apply at Autodesk in building ML-driven control systems that optimize reliability at the edge.
By 2030, connected vehicles are projected to represent over 90 percent of new car sales, up from roughly 50 percent today. This ubiquity transforms in-car delivery from experimental to baseline infrastructure. Each vehicle becomes an authenticated, networked node capable of secure data exchange, contextual policy enforcement, and machine learning inference in situ.
This shift demands systems that issue ephemeral credentials, perform real-time context verification, and generate immutable event records. These elements ensure unattended delivery remains deterministic at national scale. Predictability here is a function of consistent state management—every request, unlock, and validation recorded, replayable, and debuggable.
At Amazon, Hirani led backend and service decomposition for Key In-Car Delivery across U.S. markets, integrating with OEM APIs from General Motors and Volvo Cars for access management and session lifecycle orchestration. The public rollout across 37 cities supported about 7 million eligible vehicles, establishing cars as first-class, auditable compute endpoints for last-mile logistics.
“The right endpoint is the one already in the user’s operational loop. When the car itself becomes an authenticated address, delivery fits the system instead of breaking flow,” notes Hirani.
Systemic security determines whether connected delivery scales sustainably. In 2024 alone, over 400 new automotive cyber incidents were recorded, with more than a third involving direct system manipulation. Such trends elevate the importance of scope-limited tokens, full-chain audit logs, and OEM-aligned policy gates.
Without granular credentials and traceable event context, forensic workflows collapse and partner trust erodes. Secure systems must capture the why behind each unlock—linking actor identity, authorization scope, and environmental data.
Hirani architected the credentialing and audit subsystem for Amazon Key In-Car Delivery, implementing time-boxed, GPS-constrained tokens and tamper-evident event signing. This design prevented roughly 500,000 failed deliveries annually in covered regions, with every transaction reviewable via signed records. OEM security teams validated the cryptographic and policy frameworks, enabling sustained rollout without added manual escalation.
“Security earns the right to scale. Scoped access and auditable actions convert reliability from an assumption into an invariant,” says Hirani.
With parcel volume exceeding 22 billion units annually in the United States, delivery platforms must sustain hundreds of thousands of concurrent, time-bound sessions without degradation. Here, reliability derives from state idempotency, distributed coordination, and real-time observability.
Machine learning models augment this layer by predicting contention and dynamically adjusting session lifecycles. When session management services expose state transitions, retries become safe operations instead of failure modes.
Hirani’s Key In-Car services supported more than 100,000 simultaneous sessions at over 99.99 percent availability, ensuring token freshness and preventing concurrent access conflicts. This architecture minimized stale-state retries, stabilized route cadence, and absorbed regional spikes without manual intervention.
“Concurrency is difficult only when the state is opaque. When state is observable and lifecycles remain simple, load becomes routine,” she notes.
Approximately 58 million packages were stolen in 2024 in the United States, incurring losses exceeding 12 billion dollars. In-car delivery, secured by scoped credentials and verified telemetry, reduces this exposure by ensuring drops occur within locked, verified compartments under valid sessions.
Preventing theft upstream of delivery not only improves customer satisfaction but also reduces model retraining overhead and error propagation. Machine learning classifiers trained on delivery context can dynamically route packages toward low-risk endpoints, minimizing false unlocks and redundant attempts.
Hirani’s system returned roughly two minutes per delivery and saved an estimated 250,000 driver hours annually, largely by eliminating second-attempt loops and theft-related failures.
“The best metric is the quiet one: fewer misses, fewer claims, and consistent on-time completion,” says Hirani, an editorial board member at the Indiana Journal of Multidisciplinary Research.
As edge computing and connected logistics expand, projected at 424 billion and 259 billion dollar markets respectively by 2030, the architectural advantage belongs to systems that combine policy-aware endpoints with identity-scoped automation. These enable closed-loop ML models that not only act but also justify every action.
Hirani’s trajectory, from Amazon’s Key In-Car to Autodesk’s ML-enabled control systems, demonstrates how identity, telemetry, and explainability can align under one design principle: reliability is earned through verifiability.
“Reliability earns the right to scale. Intelligence turns that scale into measurable impact,” she concludes.