Software Engineer, Command Centre (Vision)

Laelaps AI Ltd.; Zürich; Vor Ort; CHF 110.000 - 150.000; Vollzeit

Zusammenfassung

Laelaps AI Ltd. is seeking a Software Engineer on Command Centre to own vision-agents that turn camera and robot feeds into tracked objects and decisions an operator can trust.

You will work with closed-set and open-vocabulary detectors, trackers and a vision-language model across GPU servers on customer sites and in the cloud. You will start from state-of-the-art models, optimize them for live-site performance, and own the service end-to-end—from data pipelines and models to evals, privacy and

Qualifikationen

• 3+ years building real-time computer vision or ML systems in production.
• Strong Python and software engineering: async services, testing, profiling and clean interfaces.
• Hands-on experience with object detection, multi-object tracking and re-identification across multiple cameras.
• Real-time video: RTSP, GStreamer or DeepStream, hardware decoding and multi-stream pipelines.
• Model optimization and serving: ONNX, TensorRT, quantization, and serving large models with vLLM or similar.
• Experience building with vision-language models and evaluating them rigorously.
• Docker and Kubernetes.

Aufgaben

• Service ownership: architecture, code quality, performance and behaviour on live sites for the vision-agents service.
• Streaming at scale: RTSP ingest from cameras and robots through GStreamer and DeepStream, batching, reconnects, latency, and adding or removing cameras at runtime without disturbing the others.
• Detection, tracking and re-identification: closed-set and open-vocabulary detectors, multi-object tracking, and cross-camera ReID that keeps identities through occlusions.
• Natural-language monitoring: the loop in which a vision-language model reasons over frames and tracking context to decide what is an alert, and the prompt contract with the console.
• Evaluation: fixture-based evals and benchmarks so no model, prompt or engine change ships without evidence, measured as false alerts and misses on real site data.
• Performance and GPU budget: TensorRT engines, INT8 and FP16 quantization, and scheduling many camera streams and a vision-language model on one GPU, including admission control for model traffic.
• Privacy by design: privacy zones and face pixelation that fail closed before any frame is stored or analysed.
• Access control: licence-plate reading and vehicle and person rules at site gates.
• Deployment: containers on Kubernetes, on-prem and in the cloud, together with our infrastructure engineers.

Kenntnisse

Python programming

Real-time computer vision

Software engineering

Profiling GPU pipelines

Vision-language models

Tools

Docker

Kubernetes

ONNX

TensorRT

GStreamer

DeepStream

vLLM

Jobbeschreibung

Our Mission

At Laelaps AI, we believe robotics is entering a transformative decade, much like the arrival of the internet. Advances in AI, cloud computing, and hardware are reshaping what autonomous systems can do. Our mission is to build the intelligent software that powers physical security in the real world - enabling robots and sensors to handle dangerous and critical tasks that humans shouldn't have to. By engineering the orchestration layer for intelligent security, we aim to create a world that is safer, more secure, and more resilient.

We're a strong founding team based in Zurich, backed by visionary investors and advisors. We are engineering the future of security today!

The Role

As a Software Engineer on Command Centre, you will own vision-agents: the service that turns every camera and robot feed on a site into tracked objects and decisions an operator can trust. Operators describe what to watch in plain language, such as "alert when a car enters the perimeter", and the service runs it continuously: GPU-accelerated ingest of every stream, a fast detector, tracker and re-identification at 8 FPS per camera, and a vision-language model that reasons over the scene and the tracks to decide what is worth an alert. It runs on GPU servers at customer sites and in our cloud, with dozens of cameras per machine.

This is a builder's role first: you start from the best available models (closed-set and open-vocabulary detectors, re-identification networks, open-weight vision-language models) and make them fast, correct and dependable on live sites. You own the service end to end, from the pipeline and the models to evals, privacy and deployment, and you work with our full-stack engineer on the operator console and the API between them. Perception on the robot itself sits with Robot Autonomy; everything off the robot that sees is yours.

What You'll Work On

• Service ownership: architecture, code quality, performance and behaviour on live sites for the vision-agents service.

• Streaming at scale: RTSP ingest from cameras and robots through GStreamer and DeepStream, batching, reconnects, latency, and adding or removing cameras at runtime without disturbing the others.

• Detection, tracking and re-identification: closed-set and open-vocabulary detectors, multi-object tracking, and cross-camera ReID that keeps identities through occlusions.

• Natural-language monitoring: the loop in which a vision-language model reasons over frames and tracking context to decide what is an alert, and the prompt contract with the console.

• Evaluation: fixture-based evals and benchmarks so no model, prompt or engine change ships without evidence, measured as false alerts and misses on real site data.

• Performance and GPU budget: TensorRT engines, INT8 and FP16 quantization, and scheduling many camera streams and a vision-language model on one GPU, including admission control for model traffic.

• Privacy by design: privacy zones and face pixelation that fail closed before any frame is stored or analysed.

• Access control: licence-plate reading and vehicle and person rules at site gates.

• Deployment: containers on Kubernetes, on-prem and in the cloud, together with our infrastructure engineers.

Who We're Looking For

We're looking for an engineer who has shipped a real-time computer vision system into production and kept owning it after launch. You are as comfortable profiling a GPU pipeline as choosing a tracker, and you judge a model by its false alerts on a live site at night, not by a benchmark. You will own a service that operators and other teams depend on, so you write code others can change safely and you measure before you claim.

Your Background

• 3+ years building and running computer vision or ML systems in production.

• Strong Python and solid software engineering: async services, testing, profiling and clean interfaces.

• Hands-on experience with object detection, multi-object tracking and re-identification across multiple cameras.

• Real-time video: RTSP, GStreamer or DeepStream, hardware decoding and multi-stream pipelines.

• Model optimization and serving: ONNX, TensorRT, quantization, and serving large models with vLLM or similar.

• Experience building with vision-language models and evaluating them rigorously.

• Docker and Kubernetes.

Nice to Have

• NVIDIA DeepStream or Triton Inference Server.

• Open-vocabulary detection or segmentation models.

• Licence-plate recognition or access-control systems.

• Privacy engineering for video: masking, anonymization, GDPR.

• Background in surveillance, security or defence systems, including on-prem or air-gapped deployment.

• C++ for performance-critical paths.

• Publications at top venues (ICCV, CVPR, ICLR, NeurIPS) are a plus but not expected. We value shipped systems more.

What We Offer

• Ownership: you are able to ship products and deliver project end-to-end.

• Mission: autonomous security that keeps people and critical sites safe, including in defence.

• Career path: a ground-floor seat with real runway. Prove your value and you will not have barriers to grow.

• Team: work directly with PhD-level co-founders in AI, Robotics, and Physics, alongside a strong (and fun) founding team.

• Compensation: Competitive equity/salary package

• Culture: International founding team that is serious about building but does not take itself too seriously.