Transforming fleet analytics for Novus Hi-Tech through agentic AI

Thousands of vehicles stream dashcam footage, driver alerts, GPS pings, and speed warnings across 10 countries simultaneously, and no team can manually triage every drowsiness alert or over-speeding event before it becomes an incident. Novus Hi-Tech, with over 120 enterprise customers relying on its fleet, built FleetGPT with the AWS Prototyping and AI Customer Engineering (PACE) team using the Strands Agents SDK and AWS managed AI services. AWS Architecture Blog's report records a 46 percent reduction in road-behavior violations and a 55 percent drop in drowsiness alerts per 1,000 km.

This piece is for fleet managers, customer success analysts, and Command Center operators who need different answers from the same warnings. It covers FleetGPT’s natural language querying, automated investigation, and AI-powered voice calls, which turn IoT telemetry into real-time safety interventions. It also covers the data layer, the agent model split, dashcam retrieval, and tenant-scoped queries.

01Hundreds of judgment calls per shift

For years, a drowsiness alert meant validating video footage, assessing severity across multiple signals, contacting the driver, and logging the outcome, while fixed dashboards showed the same views regardless of the question. At that scale, a company cannot staff enough operators to catch every warning in real time, and manual analysis becomes the bottleneck. The cost of a missed warning could be a preventable accident.

Novus Hi-Tech needed a system that separated a fleet manager’s question, a customer success analyst’s search, and a Command Center operator’s immediate action. Operators cycled through the same judgment calls hundreds of times per shift, while fleet managers looked for which drivers are trending toward accidents, which routes carry disproportionate risk, and which fleets are drifting on safety compliance. Operators spent too long on single events, managers could not analyze trends fast enough, and both had to switch tools or wait on data teams.

02FleetGPT analyst and intervention engine

FleetGPT replaces fixed dashboards with a natural language interface for dynamic querying, investigation, and insight generation, surfacing video evidence, driver history, and policy in a single interaction. On a critical event such as repeated drowsiness, over-speeding, or an SOS trigger, it initiates a two-way AI-powered voice call in the driver’s native language, with user approval. The system tracks acknowledgment and escalates when necessary.

The Analyst is a conversational intelligence layer where operators and fleet managers query fleet data, retrieve video evidence, and search policy documents in natural language. The intervention engine monitors incoming events, evaluates severity, and executes driver interventions with human approval. Telemetry, video, and policy documents, previously separate systems, now sit in one data layer behind that conversational interface.

03Aurora, DynamoDB, and knowledge bases

Amazon Aurora PostgreSQL is the structured analytical backbone and the primary target of the Text-to-SQL pipeline. It stores per-vehicle daily aggregates for distance, fuel, safety scores, event counts, and GPS statistics, plus live snapshots of current vehicle status. A YAML schema file with column aliases, descriptions, and worked examples goes into every SQL generation prompt, which grounds the large language model on the actual schema and reduces hallucination on column names and join paths.

Amazon DynamoDB stores the high-velocity warning stream, using low-latency reads and writes and composite keys optimized for agentic query patterns. The warning event table’s composite sort key, warning_code plus datetime, supports range queries by warning type and time window without a full table scan. A Global Secondary Index on fleet_id keeps a fleet-scoped query off other tenants’ data, and a device registry maps vehicle IDs to country metadata, the source of truth for language routing in audio alerts.

Amazon Bedrock Knowledge Bases provides managed retrieval-augmented generation over unstructured documents: standard operating procedures, compliance policies, and product documentation stored in Amazon Simple Storage Service (Amazon S3). Amazon Titan Text Embeddings V2 builds the vectors, and Amazon Aurora PostgreSQL stores them with pgvector. Operators then ask data questions and policy questions in one interface, with no context switching.

04Supervisor, SQL, and Python agents

Strands Agents assigns each agent a role and a model matched to task complexity, with a focused tool surface and a scoped system prompt that reduces hallucination and isolates failures. The Supervisor, on Claude Sonnet 4.6, decomposes intent, handles multi-step reasoning, and routes the question to a sub-agent. A SummarizingConversationManager keeps the last six messages verbatim and summarizes older turns, bounding context-window use while the conversation stays continuous.

The Text-to-SQL sub-agent, also on Claude Sonnet 4.6, uses a plan-and-execute approach. It classifies the query as vehicle data, driver behavior, or operational metrics, extracts entities, maps them to tables and columns, and writes a plan of tables and operations before it produces and executes the SQL. On an error it retries with the error message, the original plan, and the failed query in the prompt for self-correction.

Work beyond SQL, including correlation analysis, statistical profiling, and trend detection, goes to a Data Processing sub-agent on Claude Haiku 3.0. That agent generates executable Python, runs it in a Python REPL tool, and returns shaped results to the orchestrator instead of pushing that processing onto the database. Haiku 3.0 is smaller than Claude Sonnet 4.6, which keeps inference cost down, and the application is hosted on AWS Fargate for Amazon Elastic Container Service, a serverless container environment, with an Application Load Balancer as the entry point.

05Natural language dashcam retrieval

Retrieving dashcam footage used to require known vehicle IDs, a separate system, and filters applied manually. Operators now name a fleet, vehicle, warning type, date range, speed, or city in natural language. The agent turns those parameters into a DynamoDB query on the composite sort key, applies the rest as in-memory filters, and returns a paginated set of vehicle ID, timestamp, GPS coordinates, and a streamable video link in one conversation.

Multi-tenancy is enforced at the data layer for every authenticated session. Each query is scoped to the authenticated user’s authorized fleet IDs, and those IDs are injected into the agent’s session context. Video lookups, telemetry questions, and policy questions run only on fleets the signed-in user may access.