Lead Computer Vision Engineer, Counter-UAS

The Mission

Small drones have changed the threat picture. They're cheap and capable, and they increasingly show up over military installations, airports, critical infrastructure and crowds, often without broadcasting who they are. Meanwhile, our skies still hold objects no one can identify. Closing that gap is a critical national security issue, and at its core it is a computer vision challenge.

Anomaly Federal is a defense tech startup building sensor-driven detection for counter-UAS and air domain awareness. We turn cameras that are already everywhere into a detection network: mobile phones, ATAK-connected devices, PTZ cameras, and soon infrared sensors. That network finds, classifies and tracks small UAS, balloons, aircraft, and unknown or anomalous aerial objects, including UAP.

The Role

We're hiring a Lead Computer Vision Engineer to own that capability end to end, from data curation, research and model selection through a production system operators can trust in the field. You'll set the technical direction for the detection technology at the heart of our platform. You'll contribute across two interrelated areas: building and shipping production CV systems, and driving the technical research that shapes what we build next.

You'll work directly with our founder/CEO and with our Product, Engineering, Design and Data Science teams. We're looking for a first-principles thinker with an early-stage ownership mentality: decisive, proactive, and able to turn a research idea into a concrete plan a team can execute.

Think you are a good fit? Apply by sending your resume to [email protected].

The Problem You'll Solve

• Tiny, ambiguous targets. Drones may be only tens of pixels wide, sometimes far fewer, against sky, tree lines and urban clutter. Telling them apart from birds, insects and distant aircraft takes more than appearance: motion, temporal context, triangulated speed and altitude, and fusion all matter.
• Sensors you don't control. Phone and ATAK video, plus PTZ cameras that slew and zoom mid-track. Expect compression artifacts, motion blur, rolling shutter and wildly variable optics.
• Cloud and edge. 30 FPS stream-to-cloud inference, plus NVIDIA Jetson units at fixed sites and on mobile and vehicle-mounted systems.
• Fusion that matters to operators. You'll associate visual tracks with Remote ID and ADS-B so operators can deconflict cooperative aircraft and focus on what isn't broadcasting.
• What's next. Infrared is coming, and you'll shape how we add it.

Your Impact

• You own it. You'll be the company's go-to authority on computer vision, with the scope to make the calls and see them through.
• Your work leaves the lab. You'll test your models against live targets in the field. That may include range events like ISOF Range 2026, where we've participated.
• From pixels to position. Your detections don't stop at a bounding box. We fuse detections across cameras with KLV-embedded sensor metadata and triangulation to produce geo-located tracks on a map, so every gain in your models sharpens the common operating picture that operators act upon.
• Real data, hard questions. You'll work with real-world imagery that doesn't look like any benchmark. That includes objects that, by definition, don't match a training set.
• Research with a path to production. You'll help shape proposals to agencies like DARPA, NASA, the US Air Force and the US Army, then build what gets funded and take it to the field.

What You'll Do

Roadmap & Strategy

• Create the strategy for detecting, classifying and tracking aerial objects across EO video and imagery today and infrared tomorrow.
• Define and report on rigorous evaluation: COCO-style mAP and AR across IoU thresholds, with size breakouts well below COCO's "small" (<32²), down to targets only a few pixels across.
• Pair model metrics with operational ones such as precision and recall at the deployed confidence threshold, tracking quality (HOTA, IDF1), detection range, time-to-first-detection, latency and throughput.
• Measure classification at the track level, not just per frame: how quickly a track settles on the right class and how rarely it flips.
• Define operational metrics that stay meaningful across thousands of heterogeneous streams, from stable PTZ feeds to shaky, bandwidth-starved mobile video.

Engineering

• Design, train and deploy object detection and image classification models on noisy, real-world video from cameras we don't control.
• Help build our deconfliction and discrimination system, using detections, classifications and tracks to distinguish drones from birds, aircraft, helicopters and balloons, with measurable per-class accuracy and an honest "unknown" for anything that fits no known class.
• Build the data engine: ingestion, labeling, curation, hard-negative mining and continuous retraining on our growing dataset of crowdsourced video.
• Own the MLOps lifecycle: training, evaluation, deployment, monitoring and retraining.
• Integrate CV outputs into our broader detection and fusion pipeline alongside Remote ID, ADS-B, metadata triangulation and algorithmic scoring.
• Build multi-object tracking that holds up through intermittent detections, handheld shake, PTZ slew and zoom, and clutter, using camera-motion compensation so tracks follow the target, not the camera.
• Optimize models and pipelines to sustain 30 FPS per stream on Jetson and in the cloud, without giving up recall on the smallest targets, which are the first casualty of downscaling and quantization.
• Partner with other team members on system design and APIs. Raise the bar in code and architecture reviews.

Research

• Serve as a core technical contributor and writer on research grants and SBIRs to science and defense agencies.
• Track the state of the art and decide what's worth adopting. That includes foundation vision models, vision-language models, few- and zero-shot detection, small-object detection and sensor fusion.
• Make build-vs-buy-vs-fine-tune calls on models, datasets and tooling, with a clear rationale.
• Develop evaluation methods for few-pixel targets, where IoU stops being a fair yardstick.
• Turn promising research into proofs of concept that graduate into production.
• Requirements
• 5+ years building production computer vision systems. You've adopted, fine-tuned and built detection models, including substantial changes to architecture, loss or training recipe, and can show with data why each choice was right.
• Experience leading a CV or ML effort from problem definition to deployment.
• Deep, current expertise in object detection, from real-time CNN detectors (e.g., the YOLO family) to transformer-based detectors (e.g., RT-DETR), and the judgment to choose between them under a latency budget.
• Hands-on experience with multi-object tracking in video with SORT-family trackers, including using track kinematics, such as speed, trajectory and maneuver patterns to inform classification.
• Demonstrable Python, PyTorch and OpenCV experience.
• Real-time video pipelines you've built, with latency you've measured and improved.
• Ownership of the full ML lifecycle: labeling and curation, training, evaluation, deployment, and monitoring and retraining.
• Solid fundamentals in camera models, image formation and classical CV alongside deep learning. You know when a Kalman filter beats another network.
• Clear technical writing and communication, whether to an engineering peer or a federal program manager.

Nice to Have

• Edge deployment on Jetson with TensorRT, ONNX and quantization (FP16/INT8) in austere environments.
• Infrared or thermal imagery.
• Tiny-object detection in aerial or other physical-world sensor data, including its evaluation pitfalls, such as IoU's sensitivity to small localization errors on few-pixel targets and alternatives like Normalized Wasserstein Distance.
• Sensor fusion and track association with RF, radar, ADS-B, Remote ID or other metadata.
• Counter-UAS, ISR or defense experience, and familiarity with ATAK/TAK.
• Foundation vision models (e.g., DINOv2, Grounding DINO, SAM) for pretraining, auto-labeling or zero-shot detection.
• Synthetic data and domain adaptation.
• Open-set recognition (OSR), open-vocabulary detection (OVD) or out-of-distribution (OOD) detection.
• Multi-view geometry, on-the-fly multi-camera calibration methods and time synchronization.
• Track-before-detect or other point-target methods for low-SNR, few-pixel targets.
• Real-time media streaming and low-latency transport frameworks (e.g., GStreamer, FFmpeg, WebRTC, SRT, or Media over QUIC).
• C++ or CUDA.
• Publications, technical proposals or other formal research experience.
• Early-stage startup experience.
• GCP, AWS, Azure, Kubernetes, Terraform, Docker.
• An active Top Secret clearance.

Logistics

• Remote within the US, with up to 5% travel for field testing and conferences.
• Salary: $170,000–$210,000 base, plus equity.
• US citizenship / US person status required.

Equal Opportunity Employer: Anomaly Federal is an equal opportunity employer. We consider qualified applicants without regard to race, color, religion, sex, national origin, disability, veteran status, or any other characteristic protected by applicable law.