Perception Engineer

Mach Industries•Huntington Beach, CA

About The Position

Mach Industries is building an AI-forward autonomy stack for contested environments where GPS and other sensing are unavailable or unreliable. As a Perception Engineer, you will design, train, and deploy state-of-the-art vision and multi-sensor perception systems that enable navigation, targeting, and automatic target recognition on our product lines. You’ll work across deep learning, computer vision, and embedded systems to bring research-grade algorithms to real-world deployments.

Requirements

Production C++ on Linux and Python for ML/tooling; profiling, optimization, and rigorous testing discipline.
Experience diving into CUDA backends for performance optimization and debugging.
Strong with modern detection/segmentation/tracking (e.g., Retina/FCOS/DETR/Mask2D/Video models) and training/fine-tuning in PyTorch.
Proven experience building large, diverse datasets; labeling/QA pipelines; augmentation; experiment tracking; and reproducible training.
Hands-on with model compression (INT8/FP16), runtime optimization, and real-time constraints.
EO/IR imagery experience and working with real flight/test data in challenging environments.
7+ years of experience with either a BS/MS/PhD in CS/EE/Robotics or similar, or equivalent experience; track record shipping ML models to production.

Nice To Haves

Multi-modal perception experience (EO/IR + radar/LiDAR/RF) at the decision or feature level.
Robustness and safety: adversarial/rare-event testing, long-horizon reliability metrics, dataset shift/drift monitoring.
Physics-aware imaging: radiometric correction, NUC/FFC, atmospheric effects modeling; synthetic data/simulation for coverage.
MLOps and data infra: SQL/Parquet, dataset/versioning tools, CI-based validation; scalable training on multi-GPU.

Responsibilities

Build and refine detection/segmentation/tracking architectures (CNN/Transformer) for EO/IR and multi-spectral imagery; drive foundation-scale datasets, training recipes, and robust generalization to long-tail and degraded conditions.
Stand up training/eval pipelines (PR/ROC, mAP, latency, robustness suites); implement continuous regression testing and model-update loops from field data.
Optimize models for real-time embedded inference (quantization/pruning, TensorRT/ONNX Runtime), profile CPU/GPU, and meet tight throughput/latency targets on Jetson-class hardware.
Combine vision outputs with auxiliary sensing (e.g., radar/LiDAR/RF cues) for confirm/deny, association, and track management using decision-level fusion.
Create visualization, triage, and root-cause tools for rapid insight from simulation, HITL, and flight logs; define end-to-end test plans with hardware and flight teams.
Instrument health metrics, drift detection, and graceful degradation; write clear tests and documentation mapped to performance requirements.
Perform simulation-based testing with high-fidelity sensor models and validate algorithms using real-world datasets.