The US Department of Defense has quietly assembled the most ambitious artificial intelligence system ever deployed in combat operations. Known as the Maven Smart System, it now runs across every branch of the American military — Army, Navy, Air Force, Marines, and Space Force. What makes it remarkable is not just its scale, but its tech stack: a collection of open-source frameworks, commercial cloud platforms, and frontier AI models that any senior developer would recognise from their own production environments.

This is the inside story of how Maven works, who built it, and why the same tools powering your SaaS startup are now powering autonomous targeting systems on the battlefield.

From Project Maven to Maven Smart System

Project Maven started in 2017 as a modest Pentagon initiative to use machine learning for drone surveillance footage analysis. Google famously won the initial contract — and then withdrew after thousands of employees protested the company's involvement in military AI. That withdrawal created a vacuum that three companies rushed to fill: Palantir Technologies, Amazon Web Services, and Anduril Industries.

By 2025, the project had evolved from a single computer vision experiment into a full-spectrum decision support platform rebranded as the Maven Smart System. It ingests satellite imagery, signals intelligence, drone feeds, radar returns, and open-source data, then fuses everything into a unified operational picture that commanders can query in near-real time.

The Architecture: Who Built What

The Maven Smart System is not a monolithic application. It is a layered platform, and each layer has a clear owner:

Palantir provides the ontology and operating system layer. Its Foundry and Gotham platforms define how entities — people, vehicles, buildings, units — are modelled, linked, and queried. Palantir's ontology is the data backbone: it turns raw intelligence into structured, searchable objects with relationships. If you have ever worked with a knowledge graph in production, this is that concept at military scale.

AWS and Microsoft Azure supply the cloud infrastructure. Maven runs on classified cloud partitions — AWS GovCloud and Azure Government — that meet the DoD's IL5 and IL6 security requirements. These are air-gapped from commercial regions but run the same core services: compute, storage, container orchestration, and managed databases.

Anduril Industries handles the hardware edge. Its Lattice platform connects autonomous drones, ground sensors, and surveillance towers into a mesh network that feeds data back to Maven. Anduril's contribution is critical because Maven is only as good as the sensor data it receives, and Lattice ensures that data flows from the physical battlefield into the software stack with minimal latency.

The Tech Stack: Open-Source Tools at War

Beneath the proprietary platforms sits a surprisingly familiar collection of open-source technologies. Developers working on Maven are writing code against the same frameworks used in fintech, adtech, and e-commerce:

Apache Kafka handles event streaming. Every sensor reading, every drone telemetry packet, every signals intelligence intercept is published to Kafka topics. The system processes millions of events per second across distributed brokers, enabling real-time situational awareness. Kafka's durability guarantees matter enormously here — losing an intelligence event in transit could mean missing a threat.

Apache Spark runs the batch and micro-batch data processing. Raw intelligence data is cleaned, deduplicated, enriched, and aggregated through Spark pipelines before being loaded into Palantir's ontology. Spark's ability to process terabytes of imagery metadata, communications intercepts, and geospatial data makes it the workhorse of Maven's data engineering layer.

OpenCV powers the computer vision pipeline. Drone and satellite imagery passes through OpenCV-based preprocessing — object detection, image segmentation, change detection — before reaching the higher-level AI models. This is where Maven's origins in Project Maven's drone footage analysis are most visible. The models identify vehicles, structures, troop movements, and terrain changes at a pace no human analyst team could match.

Neo4j serves as the graph database for entity resolution and relationship mapping. When Maven identifies a phone number linked to a vehicle linked to a building linked to a known network, those connections are stored and traversed in Neo4j. Graph databases excel at exactly this kind of multi-hop relationship query, and in an intelligence context, discovering that a target is three hops from a known threat actor can be operationally decisive.

Open Policy Agent (OPA) enforces rules and access controls across the platform. In a system where different classification levels, national caveats, and need-to-know restrictions coexist, OPA provides a declarative policy engine. Every data query, every model inference, every user action is evaluated against OPA policies before execution. This is how Maven ensures that a NATO ally sees only the intelligence they are cleared for, or that a targeting recommendation cannot bypass required legal review steps.

The LLM Layer: From Anthropic to OpenAI

The most politically charged component of the Maven Smart System is its large language model integration. The DoD initially selected Anthropic's Claude as the primary LLM for natural language querying, report generation, and analyst assistance. Claude's reputation for safety alignment and its strong performance on reasoning benchmarks made it an attractive choice for a system where accuracy has life-or-death consequences.

That arrangement did not last. Anthropic CEO Dario Amodei publicly raised concerns about the deployment, citing the company's Responsible Scaling Policy and its commitment to ensuring AI safety. The objection forced the DoD to find an alternative, and OpenAI stepped in to fill the gap. OpenAI's models now handle Maven's natural language interface, enabling commanders and analysts to query the system in plain English rather than writing structured database queries.

The switch underscored a growing rift in the AI industry between companies willing to work with defence and intelligence agencies and those that draw a line. For developers, it also revealed something important: the LLM layer in Maven is designed to be swappable. The system uses standardised interfaces so that the underlying model can be replaced without rewriting the application logic — a pattern that every production AI system should follow.

MCP: The Glue That Connects AI Agents

Perhaps the most forward-looking component of the Maven Smart System is its adoption of the Model Context Protocol (MCP). Originally developed by Anthropic as an open standard for connecting AI models to external tools and data sources, MCP has become the protocol through which Maven's AI agents interact with the rest of the platform.

In practice, this means an AI agent running inside Maven can use MCP to query the Neo4j graph database, pull imagery from the OpenCV pipeline, request Spark jobs, or check OPA policies — all through a standardised tool-calling interface. MCP gives the AI agents structured access to the platform's capabilities without requiring bespoke integrations for each data source or service.

For the developer community, Maven's adoption of MCP is significant. It validates the protocol as production-grade infrastructure for AI agent orchestration, not just a research curiosity. If MCP can handle the latency, security, and reliability requirements of a military combat system, it can handle your enterprise workflow automation.

What Developers Should Take Away

The Maven Smart System matters beyond defence circles for several reasons:

Open-source tools scale to the hardest problems. Kafka, Spark, OpenCV, Neo4j, and OPA were not built for warfare. They were built for data engineering, computer vision, graph analytics, and policy enforcement. The fact that they underpin a system of this criticality validates the maturity and reliability of these projects.

LLM abstraction layers are essential. Maven's ability to swap from Claude to OpenAI's models without a system rewrite is a best practice that every AI-powered application should adopt. Vendor lock-in to a single model provider is a strategic risk.

MCP is becoming infrastructure. The protocol's adoption in a classified military system signals that tool-calling standards for AI agents are moving from experimental to foundational. Developers building agentic systems should be paying close attention.

Graph databases are underrated. The intelligence community's reliance on Neo4j for entity resolution and relationship discovery highlights a capability gap in most enterprise applications. If your system models complex entity relationships with relational joins, there may be a better way.

Policy-as-code is non-negotiable at scale. OPA's role in Maven demonstrates that as AI systems grow more autonomous, the rules governing their behaviour must be explicit, auditable, and programmatically enforced. This applies equally to a military targeting system and to your company's AI-powered customer service bot.

The Ethical Dimension

None of this exists in a moral vacuum. The Maven Smart System raises profound questions about autonomous weapons, algorithmic accountability, and the role of technology companies in warfare. Anthropic's withdrawal and OpenAI's engagement represent two different answers to the same question: should AI companies help build military systems?

Developers will increasingly face their own versions of this question. The tools are the same whether you are building a recommendation engine or a targeting system. The stack is the same. The difference is what sits at the end of the pipeline — a product suggestion or a kinetic strike.

Understanding what Maven is built on is the first step toward forming an informed opinion about what it should be allowed to do.