AIGP / Sectors / Defence & Security
Sector 02 — Defence & Security

Autonomy, counter-drone, authenticity.

For defence primes, critical infrastructure operators, security services and dual-use companies. AI that works in the field — from drone swarms with under-10MB edge models, through detection of AI-generated content, to perception systems for cameras in variable conditions.

Discuss a project → See case studies
What hurts the sector

Three challenges that define the next decade.

The war on NATO's eastern flank changed the calculus — what was R&D a decade ago is operational necessity today.

  1. Drone threats classical systems don't answer

    Cheap commercial drones, swarms, asymmetric use — radars and standard systems were designed for a different reality. Real-time perception, classification and decision deployed on edge devices is needed.

  2. Disinformation and deepfake as weapons

    Generative models reached a level where verifying authenticity of video and image material is no longer optional — for services, public media, institutions responsible for information stability.

  3. Staffing constraints vs operational expectations

    Linear scaling of operator headcount against the scale of threat is no longer possible. Automation of monitoring, classification and preliminary analysis tasks isn't a luxury — it's the only way to maintain capability.

Production deployments from AIGP partners

Production systems in field conditions.

The projects below run in production and are evaluated on metrics — not demo impressions.

Reinforcement learning · EdgeAutonomous drone flight

Thousands of virtual drones training in parallel, models <10MB.

Training pipeline for autonomous drone control replacing classical algorithms that require manual programming of every scenario. Deep RL algorithms (PPO) operate in continuous action spaces. Realistic simulations in PyBullet and Isaac Gym, multi-threaded simulation environments. Complex reward functions incorporate obstacle avoidance, energy minimization, and target arrival time. Models optimized to under 10MB for deployment on Jetson Nano and Raspberry Pi.

<10 MB
Edge model size
thousands
Parallel virtual agents in training
PPO
Continuous-action algorithm
Computer vision · AI-generated detectionAI-generated content detection

Pipeline on 25M+ media items, production ONNX export.

Scalable pipeline integrating multiple public datasets: 10M images plus 15M video items. ViT-L models for image classification, VideoMAE and V-JEPA for temporal video analysis. Distributed training with custom data pipeline. ONNX export for low-latency production inference.

~97%
Image classification accuracy (ViT-L)
~93%
Video accuracy (VideoMAE, V-JEPA)
25M+
Media in training pipeline
Computer vision · Field conditionsObject detection from cameras in variable conditions

Over 98% accuracy on production data, 3M+ images.

Dataset of 1M real images plus ~1M synthetic ones generated with BLIP/CLIP models. ViT-huge model trained on 8×A100 cluster with distributed training. Pipeline designed for high-volume I/O with minimized bottlenecks. Directly adaptable to area monitoring, breach detection, and vehicle-camera perception.

>98%
Accuracy on production data
3M+
Images in dataset (real + synthetic)
800k
Test set
Defence sector · conversation

A project that matters? Let's start with a conversation.

We understand the sector's rhythm. First meeting without NDA, without going into details that require formalization — just to assess whether it's a conversation for us.

Book a call →