Senior Applied ML Engineer, Computer Vision

Job Description

Company Description Have you ever looked up at the cameras as you drive past an intersection and wondered what is going on behind the scenes? Whenever you pick up a rental car, do you know what happens to the pictures and videos taken by the cameras near the exit of the parking lot? Would you like to know how NCIS Special Agents track down vehicles of interest? Come find out at ALPR Solutions.

\n

Role Description We are seeking a highly experienced computer vision expert to design, improve, optimize, and productionize advanced computer vision models and inference engines. This role sits at the intersection of deep learning research, neural network architecture, computer vision, and high-performance inference engineering. The ideal candidate is comfortable moving well beyond simply training an existing model: they can analyze network architecture at the layer and node level, experiment with model structure and activation functions, diagnose accuracy and performance issues, optimize inference pipelines, and translate research models into production-ready computer vision engines. You will work on computer vision problems involving object detection, image classification, OCR, visual embeddings, image retrieval, image similarity, attribute recognition, and multimodal / vision-language models. This position requires hands-on expertise with technologies such as PyTorch, ONNX, OpenCV, OpenVINO, CNNs, Transformers, CLIP-style models, and Vision-Language Models (VLMs).

\n

Job Responsibilities

\n

Neural Network Research & Architecture

\n

\n
• Analyze and modify neural networks at the layer, block, channel, node, and tensor level. 
\n
• Experiment with network depth, width, feature dimensions, convolutional blocks, normalization strategies, residual connections, attention mechanisms, and output heads. 
\n
• Evaluate tradeoffs between accuracy, inference latency, model size, and memory/CPU utilization.
\n
• Experiment with activation functions such as ReLU, Leaky ReLU, SiLU/Swish, GELU, and others.
\n
• Investigate the effect of input resolution, feature-map resolution, receptive field, stride, padding, kernel size, and network capacity on model performance.
\n

\n

Model Training & PyTorch Development 

\n

\n
• Modify models and implement custom layers, heads, loss functions, and training strategies. 
\n
• Analyze training behavior including convergence, overfitting, underfitting, class imbalance, hard-negative examples, and ambiguous classifications. 
\n
• Develop effective data augmentation, normalization, preprocessing, and sampling strategies. 
\n
• Perform model ablation studies to determine which changes materially improve performance. 
\n
• Diagnose model confusion between visually similar classes (such as side-profile GMC Yukon versus Chevy Tahoe) and develop techniques to improve discriminative capability.
\n

\n

OpenCV & Image Processing

\n

\n
• Develop high-performance image processing and computer vision pipelines using OpenCV. 
\n
• Implement preprocessing and post-processing including resizing, cropping, normalization, filtering, geometric transformations, color-space conversion, thresholding, image enhancement. 
\n
• Understand the effect of interpolation, compression, image format, resolution on model accuracy. 
\n
• Diagnose discrepancies caused by differences between training and production image pipelines.
\n

\n

Object Detection, Classification & OCR 

\n

\n
• Develop and improve models for: object detection, image classification, fine-grained visual classification, OCR / text recognition, attribute classification, region classification 
\n
• Work with architectures such as YOLO and other modern CNN/Transformer-based detectors. 
\n
• Research OCR architectures such as CTC-based and sequence-based recognition models.
\n
• Analyze character-level confusion (such as “O” vs. “0”) and difficult visual cases. 
\n
• Tune OCR network architecture, feature extraction, sequence length, decoding, and training data.
\n
• Develop methods for handling ambiguous characters, unusual layouts, stacked characters, low-resolution images, blur, glare, weather conditions, poor illumination, and image compression.
\n

\n

ONNX & Model Interoperability

\n

\n
• Export and convert PyTorch models to ONNX for deployment, considering ONNX computational graphs, operators, tensors, shapes, dynamic/static dimensions, and opset compatibility. 
\n
• Validate model outputs across PyTorch, ONNX Runtime, OpenVINO, C++, and other inference environments to diagnose model conversion problems and any differences in inference. 
\n
• Optimize ONNX graphs, including operator fusion, constant folding, and graph simplification.
\n
• Identify unsupported operators and redesign or replace network components when necessary.
\n

\n

Embeddings, CLIP & Image Search

\n

\n
• Evaluate embedding architectures and vector databases for visual search/similarity use cases.
\n
• Work with CLIP, SigLIP, MobileCLIP, EVA-CLIP, and related multimodal embedding architectures. 
\n
• Benchmark embedding models for retrieval accuracy, mAP, View email address on ziprecruiter.com, F1, ROC/AUC, embedding quality, NDCG, inference latency, model size, and CPU/GPU requirements.
\n
• Evaluate tradeoffs between general-purpose foundation versus domain-specific embeddings. 
\n
• Develop similarity search and nearest neighbor retrieval capable of operating at billion-plus scale.
\n

\n

Vision-Language Models & Foundation Models 

\n

\n
• Evaluate emerging Vision-Language Models (VLMs) and multimodal foundation models. 
\n
• Experiment with models such as Qwen-VL, Florence, and other sota vision architectures to evaluate whether they can supplement or replace traditional task-specific CV models. 
\n
• Develop approaches for zero-shot / few-shot classification, visual question answering, image understanding, semantic image search, automatic labeling, object and attribute extraction.
\n
• Adapt large pretrained models to ALPR use cases through fine-tuning, distillation, adapters.
\n
• Track emerging research and determine which techniques have practical production value.
\n

\n

Productionization & Computer Vision Engine Development 

\n

\n
• Work with software engineers to integrate nnet models into high-volume production systems. 
\n
• Design reusable computer vision inference components, APIs, and microservices. 
\n
• Optimize models where latency, throughput, memory/CPU constraints, and reliability are critical. 
\n
• Maintain compatibility across model versions, inference runtimes, and hardware platforms. 
\n
• Participate in model versioning, regression testing, performance benchmarking, and hypercare.
\n

\n

Desired Attributes

\n

\n
• Eye for design, care for security, and optimization for performance
\n
• Self-starter with the demonstrated aptitude to be self-sufficient 
\n
• Proven ability to effectively deliver on objectives 
\n
• Works well under the pressure of a deadline and prides him/herself on delivering on time
\n
• Strong desire to be successful and confident in one’s own success
\n

\n

Hybrid Work and Flexibility

\n

On average, the ALPR Solutions team works on-site 3 days each week at our office at 12610 Race Track Road, Tampa, Florida 33626. We are interviewing candidates on-site every Wednesday. Candidates must be located within daily commuting distance to Tampa, Florida, or be ready to relocate at their expense. We prioritize candidates with high volume live streaming data experience.