Project: AIMRIS Hub is an edge-first multimodal industrial monitoring platform for mining and construction. We utilize Vision-Language Models (VLM) and Computer Vision to automate safety and efficiency controls at quarries, construction sites, and factories.

The Core: We deploy VLMs (2-7B) and CV models on Edge devices (NVIDIA Jetson Orin) in environments where cloud access is unavailable.

The Challenge: Fitting top-tier open-source models into 16GB VRAM on Edge hardware while ensuring <2 second latency and high stability in harsh industrial conditions.

TECH STACK

• Hardware: NVIDIA Jetson Orin NX/AGX (ARM64).
• Models: YOLOv8/v10 (TensorRT), VLM (Qwen2.5-VL, Phi-3.5, Moondream2).
• Optimization: TensorRT, Quantization (AWQ/GPTQ), vLLM, ONNX.
• Pipeline: GStreamer, DeepStream, Python, Docker.

Responsibilities:

• VLM on Edge: Porting and quantizing (4-bit) heavy models. Achieving a balance between accuracy and speed (<2s latency).
• Dual-Pipeline Architecture: Implementing an asynchronous link where a fast detector (YOLO) triggers intelligent analysis (VLM). Managing GPU load balancing.
• Video Pipeline: Processing up to 8 RTSP streams without memory leaks or latency (Hardware decode -> Inference).
• Production Deployment: Building lightweight Docker images for ARM64, configuring Watchdogs, and implementing auto-recovery systems.

Requirements:

• 2+ years of experience in ML Engineering (focus on Inference & Deployment).
• CV & Object Detection: Deep understanding of YOLO, tracking metrics, and video preprocessing.
• Optimization: Hands-on experience with TensorRT, ONNX, and quantization methods (Int8/FP16/AWQ).
• Python Stack: PyTorch, OpenCV, NumPy, Asyncio, Pandas.

Nice to Have:

• Master’s or Ph.D. degree in a relevant field. 
• Experience with search technologies such as Elasticsearch or Vespa. 
• Familiarity with Generative AI techniques, including LLM fine-tuning and prompt engineering, W&B, neptune.ai, ML Flow, Vertex AI or similar. 
• Experience with knowledge graphs and semantic technologies.