--- title: "Defence and national security AI data and evaluation" url: https://ypai.ai/solutions/defence-national-security/ description: "EO/IR, geospatial, robotics, sensor, speech and evaluation data for defence and national-security AI, from specialist panels under European jurisdiction." source: "src/copy/routes (route /solutions/defence-national-security/)" --- # Specialist data and evaluation for defence and national-security AI. > EO/IR, geospatial, robotics, sensor, speech and evaluation data for defence and national-security AI, from specialist panels under European jurisdiction. EO/IR imagery, remote sensing, autonomy, sensor fusion, speech and expert evaluation. Produced by specialist panels assembled for your project, reviewed until it meets your acceptance line, and delivered from Norway. Automated checks catch what a machine can catch. A specialist rules on the rest. Proposed by this page: {n} warm sources read from the photograph ## Six data families. One delivery system. Every project enters through the data it needs. Choose a family to see the tasks, the modalities and the specialists behind it, or follow one of five worked paths through the same system. Every project enters through the data it needs. Choose a family to see the tasks, the modalities and the specialists behind it. Working with data not listed here? Name the sensor and the task in the brief. ### EO/IR, image and video - What the camera and the thermal sensor recorded, made trainable frame by frame. ### Geospatial and remote sensing - Aerial and satellite imagery labelled by people who read terrain for a living. ### Robotics and autonomous systems - Perception and behaviour data for systems that move on their own. - Outdoor navigation capture, stereo and RTK GPS ### Sensor and multimodal - Every sensor on one clock, calibrated, aligned and checked before it is labelled. - Radar, fused with camera and LiDAR ### Speech, audio and language - Speech and text in the languages, dialects and noise your system will meet. ### Expert evaluation - Specialist judgement on what a model produced, recorded against a rubric. - Thermal video, 30 fps, moving camera - Boxes on small targets, tracked across frames - Dual review on hard frames, HOTA and IDF1 per delivery - EEA processing, YPAI-managed or in your environment - Listener consensus, accuracy on sampled output - Hybrid, imagery stays in your environment - Stereo video, RTK GPS, synchronised sensors - Sensor QA before labelling, reviewer sign-off - Model output against a written rubric - In your environment, results to your team ## The panel is built for the project. We source, qualify, calibrate and operate specialist panels to your specification. The people who label a thermal frame, a satellite tile or a dialect transcript are chosen for that work and tested on your data before production starts. ### Source - Profiles drawn from our network and specialist channels to the project's specification. ### Qualify - A test on your own sample. Only those who pass it work on the project. ### Calibrate - Gold sets and agreement scoring with Cohen's and Fleiss' kappa, before and during production. ### Operate - Production, dual review, adjudication and rework, with every decision logged. ### Air and sensing ### Ground and geospatial ### Robotics and autonomy ### Maritime ### Language ### Engineering and review Network reach. Every panel is qualified against your specification before production. ### Could you work this project? Twelve crops of the frame above at native pixels. Mark the ones that hold the contact. {agree} of 12 agree with the gold set · kappa {kappa} The crops, the gold set and the threshold are this page's own. A project's test is built from your sample. ## When the object is only a few pixels, annotation becomes a review problem. A target of a dozen pixels in a moving frame is easy to miss and easy to box wrongly. The work is done at full resolution, checked across frames and ruled on by a reviewer before it counts. ### Raw frame - The sensor's frame as recorded. The target is a few pixels against the sky. ### Full-resolution zoom - The annotator works at the native pixel grid, not the preview. ### Frame-level box - A tight box on the target, drawn to the project's box rules. ### Temporal consistency - The box is checked against the frames before and after it. A jump is flagged. ### Reviewer decision - A second person accepts the frame or returns it with a reason. ### Accepted - The frame enters the dataset with its annotator, reviewer and version. Logged with annotator, reviewer and dataset version. Every one of them is drawn and checked at full resolution. The frame, the target and the readings are drawn by this page. Class list, box rules and acceptance thresholds are set in the project specification. The contact, its box and the readings are measured by this page from the frame's own pixels. Class list, box rules and acceptance thresholds are set in the project specification. ## Eight stations. One of them is held by a person. Every project runs the same line. Automated checks catch what a machine can catch. A specialist rules on the rest, and nothing is accepted past that point without a decision. Every project runs the same line. Automated checks catch what a machine can catch. A specialist rules on the rest, and that decision is what accepts an item. ### Specification - Classes, rules, formats, volumes and the acceptance criteria, written with you. ### Qualification - Sources and specialists tested against the specification. ### Collection and annotation - Captured or labelled by the qualified panel, model-assisted where it helps. ### Automated QC - Format, completeness, sensor timing and consistency checks on every item. ### Specialist review - A qualified reviewer rules on each item or sample the plan defines. ### Adjudication and rework - Disagreements resolved, returned items fixed and checked again. ### Accepted dataset - Versioned, with the evidence for why each item passed. ### Controlled delivery - Delivered into the environment the project specifies, with a manifest. ## What the sensors record, and what we do with it. - **Thermal and EO imagery and video**: Boxes, tracks, segmentation, small-target review - **Aerial and satellite imagery**: Geospatial annotation, human-only aerial segmentation - **LiDAR and point clouds**: 3D boxes, segmentation, tracking, calibration checks - **Radar**: Fused with camera and LiDAR, cross-sensor projection - **Depth and IMU**: Capture, timestamp alignment, sensor QA - **Stereo video and RTK GPS**: Outdoor navigation capture for ground autonomy - **Speech and audio**: 48 kHz / 24-bit capture, transcription, diarisation, sound events - **Text and documents**: Extraction, classification, terminology, translation - **Synchronised multimodal**: One clock, calibration data, multi-sensor alignment ## Your data runs where your rules say it runs. Three delivery models on one method. Choose where the data is stored, where the work is done and where the result is handed over. ### YPAI-managed - We run the environment. Data stays in the EEA by default. ### In your environment - The panel works inside the platform you provide. The data stays where it is. ### Hybrid - Each part sits where the project needs it, agreed at scoping. - **Jurisdiction**: Norway · GDPR - **Processing**: EEA by default - **Contributors**: Identity-verified - **Isolation**: Data isolation, client-specific environment - **Control**: Change control, versioned manifests - **Erasure**: 30-day end-of-contract - **Contracts**: Standard DPA terms · SCCs available - [Data residency in the EEA](https://ypai.ai/data-residency-eea/) - [Private and EEA deployment](https://ypai.ai/ai-implementation/private-enterprise-deployment/) ## Specialist judgement, measured. Some outputs only an expert can grade. We design the rubric with you, put qualified reviewers on it and record every judgement so the result can be audited. ### Rubric and benchmark design - What good looks like, written down before anyone grades. ### Dual review and adjudication - Two independent judgements, and a third where they disagree. ### Expert grading of model output - Engineers and domain experts score what the model produced. ### Red teaming - People trying to make the system fail, with every attempt logged. ### Robustness under domain shift - Degradation measured across terrain, weather, lighting, device and acoustics. ### ASR and TTS evaluation - Recognition and synthesis measured on your languages, dialects and noise. ### Media integrity - Real and generated pairs across face, body and voice for detection models. When the data has to run inside a system, the same team builds it. ## When the case is rare, the data already exists, or it has to run inside a system. The team that labels and reviews also generates the cases you cannot collect, licenses inventory that already exists and builds the system the data feeds. ### Synthetic and simulated data Rare events and edge cases generated against a target distribution you define, then checked against real data before they reach training. The drawn crops are generated in this tab from the sky and the contact measured on the thermal frame above, then measured back the same way. A project's generator is built against your target distribution and validation method. ### Dataset sourcing and licensing License what already exists and collect only what is missing. - Inventory found, sampled and technically reviewed - Exclusive or non-exclusive terms, by use case and geography ### AI systems in your environment When the data has to run inside a system, the same team builds it where you control it. - Sovereign deployment in your own environment - Document agents with review queues and approval gates ## Start with one modality, one task and one acceptance line. Tell us what the system has to see or hear, and under which rules. We reply with a feasibility read and a pilot plan against your own acceptance criteria. A bounded pilot carries its own acceptance criteria. You judge it against your own data. - A feasibility read on your data - A pilot plan against your own acceptance criteria Name the modality, the task and the rules it runs under. We come back with a feasibility read on your data and a pilot plan. Enquiry details are treated as confidential. Every accepted item carries a person's decision.