Smart Arbitrage Technologies

Middle Computer Vision & ML Engineer — Founding Team

About Us

We are a newly funded MilTech startup building an on-premises platform for automated annotation of petabytes of UAV video and training computer vision models for autonomous flight. We have a strategic partnership with the largest Ukrainian UAV manufacturer and access to what is the largest privately owned UAV video dataset in the world. We are at the very beginning — the technical plan exists, the funding is secured, the data is real — and now we need the right people to turn it into reality.

The Role

We are looking for a Middle CV/ML Engineer to join as one of the first engineers on the team. You will work closely with the technical lead to build the ML pipeline from the ground up — implementing core components across data processing, model training, and deployment.

This is an early-stage startup, so the tech stack and processes are still being shaped. You will have real input into the tools and approaches we adopt, and your work will directly define how the platform operates. We are looking for someone who is hands-on, curious, and comfortable building things that don't exist yet.

What You Will Work On

Data Pipeline & Preprocessing
- Build video preprocessing pipelines: clip segmentation, frame extraction, scene detection, quality filtering
- Implement embedding extraction and data curation workflows for selecting diverse training subsets from a massive video corpus
- Prepare and export datasets in standard formats for training and annotation

Model Training & Experimentation
- Train and evaluate computer vision models for object detection, classification, tracking, and segmentation
- Run experiments, track results, and iterate on model architectures and hyperparameters
- Implement data augmentation strategies for robustness across weather, lighting, and occlusion conditions
- Contribute to distributed multi-GPU training pipelines

Active Learning & Annotation Support
- Implement auto-labeling pipelines with confidence-based routing
- Build tooling around the annotation platform (pre-annotation, format conversion, quality validation)
- Maintain automated quality checks for label consistency

Edge Deployment
- Export and optimize models for inference on NVIDIA Jetson (ONNX, TensorRT, FP16/INT8 quantization)
- Run accuracy and latency benchmarks on edge devices
- Help maintain the model delivery pipeline from training to deployment

What We Are Looking For

Must Have
- 3+ years of hands-on experience in computer vision and deep learning
- Solid PyTorch skills — comfortable writing training loops, custom datasets, and model code
- Experience with at least one detection or segmentation framework (YOLO family, Detectron2, MMDetection, or similar)
- Understanding of model evaluation metrics (mAP, IoU, precision/recall)
- Familiarity with experiment tracking (MLflow, W&B, or similar)
- Python as primary language; clean, well-structured code
- Comfortable working in Linux and containerized environments (Docker)

Nice to Have
- Model optimization experience (ONNX export, TensorRT, quantization)
- NVIDIA Jetson or other edge/mobile deployment
- Multi-object tracking (ByteTrack, DeepSORT, or similar)
- Experience with video processing pipelines (FFmpeg, OpenCV, decord)
- Data curation or annotation tools (FiftyOne, CVAT, Label Studio)
- Distributed training (DDP, Ray Train)
- UAV / drone / robotics or MilTech domain experience

Proposed Tech Stack

This is our starting point — the stack is still being finalized, and you will have input into the choices we make.

- Training: PyTorch, distributed training (DDP/FSDP) 
- Data processing: Ray Data or equivalent scalable framework 
- Data curation: FiftyOne, embedding-based selection 
- Annotation: CVAT with pre-annotation and active learning 
- Experiment tracking: MLflow 
- Edge: ONNX → TensorRT → NVIDIA Jetson 
- Infrastructure: Kubernetes, on-prem GPU cluster 
- Storage: S3-compatible object storage (Ceph) 

What We Offer

- Join an early-stage team where your contributions shape the product from day one
- Work across the full ML pipeline — not just one narrow piece
- The largest privately owned UAV video dataset in the world — petabyte-scale, impossible to replicate
- Dedicated on-prem GPU resources — no cloud queues or budget negotiations for compute
- Open-source-first approach with no vendor lock-in
- Grow fast in a small team with direct mentorship from the technical lead
- Competitive salary
- Stock option
 

Lead Computer Vision & ML Engineer — Founding Role

About Us

We are a newly funded MilTech startup building an on-premises platform for automated annotation of petabytes of UAV video and training computer vision models for autonomous flight. We have a strategic partnership with the largest Ukrainian UAV manufacturer and access to what is the largest privately owned UAV video dataset in the world. We are at the very beginning — the technical plan exists, the funding is secured, the data is real — and now we need the right people to turn it into reality.

The Role

We are looking for a Lead CV/ML Engineer as a founding member of the engineering team. This is not a role where you inherit a codebase and maintain it. You will define the ML architecture, choose the tools, build the pipeline from scratch, and grow a team around it as the platform matures.

The technical plan outlines the direction — petabyte-scale video processing, active learning, edge deployment on Jetson — but the specific implementation decisions are yours to make. We value strong technical judgement over checkbox experience with a predefined stack.

What You Will Own

ML Architecture & Technical Direction
- Define the end-to-end ML pipeline: from raw video ingestion through data curation, training, and deployment to UAVs
- Evaluate and select the tools and frameworks that best fit the problem (our current thinking: PyTorch, Ray, MLflow, FiftyOne — but nothing is set in stone)
- Establish training infrastructure, experiment tracking, and model versioning from day one
- Make build-vs-integrate decisions across the stack

Data Pipeline & Curation
- Design video preprocessing pipelines: clip segmentation, frame extraction, scene detection, embedding extraction
- Build a data curation workflow to select diverse, high-quality training subsets from a massive raw corpus
- Define the annotation ontology and labelling policy for detection, classification, and multi-object tracking

Model Development
- Train and iterate on computer vision models for object detection, scene classification, tracking, and segmentation
- Implement distributed multi-GPU training for large-scale experiments
- Design the active learning loop: auto-labelling with confidence routing, human review integration, retrain triggers
- Define quality control metrics and model promotion criteria

Edge Deployment
- Optimize models for real-time inference on NVIDIA Jetson (ONNX, TensorRT, INT8 quantization)
- Validate edge accuracy against training baselines
- Design the model delivery pipeline from registry to deployed UAVs

Team Building
- Hire and mentor ML engineers as the team grows
- Establish engineering practices: code review, testing, documentation
- Own the ML technical roadmap and communicate it across the company

What We Are Looking For

Must Have
- 5+ years in computer vision and deep learning, with production deployments
- Strong PyTorch proficiency, including distributed training
- Experience with object detection and tracking pipelines (YOLO family, RT-DETR, Detectron2, or similar)
- Hands-on model optimization for edge or mobile deployment (TensorRT, ONNX, SNPE, or similar)
- Experience building training infrastructure (pipelines, experiment tracking, CI/CD for ML)
- Proven ability to lead a small engineering team (hiring, mentoring, technical guidance)
- Python as primary language; clean, production-grade code

Strong Plus
- UAV / drone / robotics domain experience
- NVIDIA Jetson deployment experience
- High-performance video processing pipelines (DeepStream or custom)
- MilTech or defense-related work
- Experience defining ML architecture from scratch at an early-stage company
- C++ for performance-critical components

Proposed Tech Stack

This is our starting point, not a mandate. The right candidate will validate, adjust, or replace these choices based on their experience.

- Training: PyTorch, distributed training (DDP/FSDP) 
- Data processing: Ray Data or equivalent scalable framework 
- Data curation: FiftyOne, embedding-based selection 
- Annotation: CVAT with pre-annotation and active learning 
- Experiment tracking: MLflow 
- Edge: ONNX → TensorRT → NVIDIA Jetson 
- Infrastructure: Kubernetes, on-prem GPU cluster 
- Storage: S3-compatible object storage (Ceph) 

What We Offer

- A founding engineering role with direct influence on the technical direction of the company
- Ownership of the full ML pipeline end-to-end — from raw video to models flying on UAVs
- The largest privately owned UAV video dataset in the world — petabyte-scale, impossible to replicate
- Dedicated on-prem GPU resources — no cloud queues or budget negotiations for compute
- Open-source-first approach with no vendor lock-in
- The opportunity to build and lead an ML team from the ground up
- Competitive salary
- Stock option