Senior Computer Vision Engineer - Core Perception

About Aegis

Aegis builds computer vision for racquet sports. Our platform, Aegis Vision, officiates and analyzes play in real time: automated line calls, live scoring, and analytics, with replay and broadcast-quality output on top.

Pickleball is the fastest-growing sport in the US, and facilities have no good answer for disputed calls, officiating cost, or the content players want after a match. We sell to facilities, and the system is deployed and in use today.

We are a small, fast team with real hardware in the field. The computer vision core is the product. That is what this role owns.

The role

You will own the core perception pipeline: the models, the multi-camera geometry, and the real-time system that turns video into calls a player trusts. You will work directly with the CTO and founders, with a few engineers around you on the app, analytics, and infrastructure.

This is a builder role with architecture responsibility. The system is live. We need someone who can keep raising its accuracy, speed, and robustness, and who can also define how the pipeline is structured so it holds up as deployments scale. You will not be handed a spec; you will help write it.

What you'll work on

• Detection and tracking of a small, fast object. At pro speeds a 2.9-inch ball is a few pixels wide in 4K, streaks between frames, and disappears behind players, paddles, and the net. Track it across multiple synchronized streams under daylight, LED, and mixed lighting.
• Multi-camera calibration and 3D reconstruction. Recover the court, the cameras, and the ball's path in 3D, and keep it accurate across facilities where every mounting point, ceiling, lens, and floor is different.
• Event detection and game state. A line call is whether a 2.9-inch ball touched a 2-inch line. Beyond that: bounce vs. volley vs. net cord, who hit it, serve sequences, side-outs, and the rest of a rulebook that is really a state machine with edge cases.
• Player tracking and pose. Where each player is and what they are doing, frame by frame, as input to both officiating and analytics.
• Real-time inference. Several 4K streams per court, a fixed GPU budget, and a latency target tight enough to make the call before the players start arguing about it.
• Pipeline architecture. Own the structure from ingest and sync through perception, game state, and outputs: module boundaries, interfaces, testing, observability, and how we deploy and update it as the footprint grows.
• Ground truth and evaluation. Build evaluation sets with ground truth good enough to grade a referee, measure accuracy rigorously, and drive it toward the standard pro tours expect.
• Model lifecycle. Training-data pipelines, labeling, and retraining as new facilities, surfaces, and lighting conditions come online.
What we're looking for

• 5+ years building computer vision systems that shipped to production, not only research or notebooks.
• Hands-on multi-view geometry: camera calibration, epipolar geometry, triangulation, done on real cameras you had to debug.
• Object detection and tracking in video, with the judgment to pick the right model for the latency budget rather than the biggest one.
• Real-time inference on GPUs: TensorRT, ONNX Runtime, or CUDA, with hard latency constraints.
• Strong Python and C++; PyTorch (or equivalent); OpenCV.
• You have designed or substantially shaped the architecture of a perception pipeline: modules, data flow, interfaces, and how it is tested, not only the models inside it.
• You measure. You build the eval before you tune the model, and you can say what the accuracy is and why.
• Comfortable in a small team, owning a problem end to end, and shipping to physical installs where the failure mode is a wrong call in front of players.
Nice to have

• Sports tracking, officiating, or broadcast analytics systems.
• Industrial / machine vision cameras and multi-camera synchronization.
• Deploying and operating ML systems in production, including remote monitoring and updates.
• Human pose estimation.
• Video pipeline tooling: GStreamer, FFmpeg.
• Early-stage startup experience.
• You play pickleball or tennis and have argued a line call.
Why Aegis

• The problem is hard and the feedback is immediate. Your work makes the call live, in front of players. You will know quickly whether it works.
• You own the core. Small team, direct line to the founders and CTO, and the perception pipeline is yours to shape.
• It is real. The system is deployed and in use, and a fast-growing sport is pulling us forward.