Handling telemetry data from thousands of vehicles across multiple countries challenged Novus Hi-Tech’s ability to analyze, act, and automate safety insights at scale. Collaborating with AWS PACE, they implemented FleetGPT—a multi-agent AI system unifying telemetry, video, and policy data to empower teams through natural language queries and AI-driven interventions.
- AI-driven real-time safety calls reduce incidents by nearly half
- Unified data layers enable faster, contextual multi-role insights
- Multi-database design enhances query efficiency and event processing
Infrastructure signal
Novus Hi-Tech’s infrastructure integrates multiple AWS managed databases to meet distinct operational needs. Amazon Aurora PostgreSQL handles structured, analytical workloads storing daily vehicle aggregates and live statuses, supporting complex text-to-SQL transformations to minimize AI query errors. Meanwhile, DynamoDB efficiently processes high-velocity warning event streams with composite keys and global secondary indexes, optimized for filtered, time-ranged event retrieval without costly scans.
This design balances cost and reliability by distributing storage and query loads between relational and NoSQL stores. It also facilitates a unified data layer accessible by the AI agents in real-time. Embedding detailed schema context into prompts helps the large language models reduce hallucinations and align their outputs tightly with actual database structure, improving reliability in automated insights and interventions.
Developer impact
Developers gain a modular, AI-first architecture replacing static dashboards with a conversational interface capable of natural language querying and proactive intervention workflows. The FleetGPT multi-agent framework enables developers to focus on optimizing agent behaviors and query grounding instead of building manual dashboards or alert pipelines. This reduces development complexity and accelerates feature iterations.
The AI-powered automation, including two-way driver voice calls with escalation logic, significantly cuts manual operator workload and operational overhead. Developers must now maintain AI prompt engineering, agent orchestration, and integration of heterogeneous data streams instead of traditional event monitoring systems, shifting workflows towards applied AI integration and system-wide observability enhancement.
What teams should watch
Operations, safety, and product teams should monitor metrics around AI call success rates, event acknowledgement latency, and overall reduction in road incidents to validate AI intervention impact. Observability pipelines must extend beyond raw telemetry to include AI agent decisions and voice interaction logs for compliance and continuous improvement.
Development teams should prioritize maintaining schema synchronization and AI prompt accuracy to prevent hallucinations and operational errors. Monitoring database performance and query latency across Aurora and DynamoDB layers will be critical for sustaining real-time responsiveness. Finally, cross-team collaboration between AI, infrastructure, and domain experts should focus on expanding FleetGPT capabilities to cover new fleets, event types, and regional languages as the system scales globally.