01Overview
Role & responsibilities
We are looking for a Sr. AI Scientist to own and scale our Vision AI Center of
Excellence (COE). You arent joining a research lab to write theoretical papers; you are here
to build production-grade, real-time spatial analytics that drive immediate revenue in heavy
industries like ports, logistics, and manufacturing.
You will take our working Ubuntu CPU-based Edge relay software (YOLO + RTSP/RTMP
encoding) and turn it into an enterprise-grade live video analytics platform. You will work
alongside our Platform team to deploy models directly into turnkey Edge NVR hardware
boxes and air-gapped environments.
Core Responsibilities
Spatio-Temporal Pipeline Engineering: Advance our current YOLO pipeline from
single-frame object detection to multi-object tracking (MOT) across space and time
for vehicle tracking, worker safety compliance, and people analytics.
VLM Fine-Tuning & Quantization: Adapt, fine-tune, and optimize state-of-the-art
open-weight Vision-Language Models (e.g., Qwen2.5-VL/Qwen3-VL, LLaVA, SAM) for
highly localized, industry-specific tasks.
Edge Optimization: Work closely with MLOps to compress models using quantization
frameworks (AWQ, GPTQ) so complex tracking and safety logic can run efficiently on
NVR boxes and CPU/NPU edge nodes.
Synthetic Data Pipelines: Build automated data curation and synthetic data
generation loops to handle poor lighting, rusted container codes, and unique
industrial edge cases.
Required Technical Skillset
Experience: 35 years of hands-on experience deploying computer vision models
into real-world production settings.
Frameworks: Deep expertise in PyTorch, OpenCV, and the Hugging Face ecosystem.
Tracking & Detection: Proven experience with YOLO variants, combined with
tracking algorithms like ByteTrack, DeepSORT, or StrongSORT.
Quantization & Serving: Familiarity with Triton Inference Server, vLLM, TensorRT-
LLM, and ONNX Runtime.
Infrastructure: Comfortable working in Ubuntu environments, handling RTSP/RTMP
video streams, and collaborating within Docker/containerized workflows. .
