Aqua IT Springfield, US
Full-time
About the Role
Description of Services/Responsibilities:
- Design and execute fine-tuning pipelines for Vision-Language Models (VLMs) on domain-specific imagery datasets, including data preprocessing, training orchestration, and hyperparameter optimization
- Develop and implement evaluation frameworks for multimodal model performance, including task-specific metrics for image understanding, visual question answering, and spatial reasoning
- Build scalable training infrastructure on AWS (SageMaker, EC2 GPU instances) for distributed fine-tuning of large multimodal models
- Engineer data pipelines for curating, annotating, and transforming geospatial imagery datasets into model-ready formats for supervised and instruction-tuning workflows
- Collaborate with applied scientists and solutions architects to iterate on model architectures, adapter strategies (LoRA/QLoRA), and inference optimization techniques
Basic Requirements
- TS/SCI with CI Poly required with current NGA eligibility and SBU/SECNet/COE accounts
- Must be willing to work in SCIF daily or as needed
- 5+ years of professional machine learning engineering experience with a focus on deep learning
- 1+ years of hands-on experience fine-tuning large foundation models (LLMs or VLMs)
- Experience with parameter-efficient fine-tuning methods (LoRA, QLoRA, adapters)
- Familiarity with supervised fine-tuning, instruction tuning, and RLHF/DPO alignment techniques
- 4+ years of advanced Python development for ML workloads
- Strong proficiency with PyTorch and the HuggingFace ecosystem (Transformers, PEFT, Datasets, Accelerate)
- Experience with distributed training frameworks (DeepSpeed, FSDP, or Megatron)
- 3+ years of experience with computer vision or multimodal models
- Understanding of vision transformer architectures (ViT, CLIP, LLaVA-family models, or similar)
- Experience processing and augmenting image datasets at scale
- 3+ years of experience with AWS ML infrastructureSageMaker Training jobs, Processing jobs, and endpoint deploymentGPU instance selection, multi-node training, and cost optimization on EC2 (P4/P5/G5/G6e)S3 data management for large-scale training datasets
- 2+ years of experience building ML evaluation pipelinesAutomated benchmarking, metric computation, and result analysisExperience with both quantitative metrics and qualitative/human evaluation approaches
- Strong software engineering fundamentals (version control, testing, CI/CD for ML workflows)
Preferred Qualifications
- 2+ years of experience with geospatial or remote sensing imagery
- Familiarity with electro-optical and SAR satellite imagery formats and characteristics
- Understanding of geospatial metadata, coordinate systems, and imagery preprocessing
- Experience with model quantization and inference optimization (vLLM, TensorRT, ONNX)
- Experience with MLOps and experiment tracking tools (MLflow, Weights & Biases, SageMaker Experiments)
- Familiarity with data annotation platforms and active learning workflows for imagery
- Experience with containerized ML workflows (Docker, ECR, ECS/EKS)
- 2+ years of experience with Authority to Operate (ATO) processes in government environments
- Implementation of NIST 800-53 controls and security compliance for ML systems
- Experience deploying models in air-gapped or disconnected environments
- Familiarity with multimodal evaluation benchmarks (MMMU, MMBench, GQA, or domain-specific equivalents)
- Publications or demonstrated contributions in computer vision, VLMs, or multimodal AI
- Experience with synthetic data generation for training data augmentation
Complete items below line after a partner is selected
#J-18808-Ljbffr