AI Inference Engineer - Model Optimization & Deployment

Responsibilities

• Optimize large-scale models (LLMs, VLMs) using advanced quantization (PTQ, QAT), mixed-precision inference workflows, and parameter-efficient fine-tuning (LoRA, QLoRA).
• Architect and implement model conversion and compilation pipelines using TensorRT and TensorRT-LLM for edge deployment.
• Perform rigorous parity checking, accuracy recovery, and latency benchmarking between PyTorch frameworks and compiled edge binaries.
• Write and optimize custom CUDA kernels and TensorRT Plugins to maximize memory bandwidth and minimize latency on AI accelerators.
• Write production-level, highly concurrent, and memory-safe C++ and Python code for real-time inference on vehicle SOCs.

Qualifications

• Deep expertise in model quantization (PTQ, QAT) and mixed-precision inference workflows (INT8, FP8, INT4, BF16/FP16).
• Proven experience optimizing large-scale models (LLMs, VLMs, or VLAs) utilizing KV-cache optimization (e.g., PagedAttention), Speculative Decoding, and Efficient Attention mechanisms (FlashAttention, Linear Attention).
• Extensive experience with model conversion/compilation pipelines (TensorRT, TensorRT-LLM) and performing rigorous parity/latency benchmarking.
• Proficiency in low-level programming for AI accelerators, specifically writing and optimizing custom CUDA kernels and TensorRT Plugins.
• Production-level C++ (14/17/20) and Python programming skills, with experience writing concurrent, memory-safe, real-time inference code for edge devices.

Bonus Qualifications

• Experience with distributed training pipelines and model/tensor parallelism (PyTorch Distributed, Ray, DeepSpeed, Megatron-LM) and runtime efficiency optimization for GPU clusters.
• Familiarity with autonomous driving perception stacks (temporal 3D object detection, BEV, 3D Occupancy Networks) and processing multi-modal sensor streams (Vision, LiDAR, Radar).
• Understanding of end-to-end autonomous driving paradigms (VLA models, closed-loop simulation validation).