Consultant, AI Engineer

Required Qualifications and Experience

• Education: B.S. or M.S. in Computer Science, Artificial Intelligence, Machine Learning, or a related quantitative field.
  

• Experience: 7+ years of experience in Machine Learning Engineering, with at least 1-2 years specifically focused on Computer Vision, Document AI, or Multimodal Large Language Models.
  

• Core Frameworks: Deep expertise in PyTorch and the Hugging Face ecosystem (Transformers, PEFT).
  

• Inference Engines: Hands-on, production-level experience deploying models using vLLM.
  

• Domain Expertise: Proven experience working with Document AI, Optical Character Recognition (OCR), Handwriting Recognition (HTR), or Vision-Language models.
  

• Image Processing: Proficiency in computer vision libraries (OpenCV, Pillow) and experience handling real-world, variable-quality mobile images, including tiling and chunking strategies.
  

• Design and optimize AI pipelines for complex document understanding. Focus on extracting structured data from mobile-captured HMIS forms, specifically tackling challenges like handwriting recognition, complex table extraction, and multilingual parsing.
  

• Research, benchmark, and fine-tune state-of-the-art Vision-Language models (e.g., Qwen-VL) and foundational OCR models on domain-specific datasets. Utilize advanced techniques (LoRA/QLoRA, DeepSpeed) to maximize accuracy on noisy, real-world mobile images.
  

• Architect and deploy production-grade inference pipelines using vLLM or similar engines. Optimize continuous batching, KV cache management, and quantization to maximize throughput while strictly maintaining our low per-page processing cost targets.
  

• Design architecture for both self-hosted/local cloud environments (like Linode) and on-premise hardware, keeping data sovereignty and cost efficiency in mind.
  

• Tune AI models for visual data optimization. Develop strategies for image chunking, tiling, and preprocessing to allow models to efficiently process high-resolution images and large, complex tables without losing context.
  

• Evaluate, select, and provision optimal cloud and on-prem GPU infrastructure to handle a target volume of 10 million forms.
  

• Assess next-generation hardware (e.g., NVIDIA Blackwell nodes) to balance massive scalability, performance, and budget efficiency.
  

• Lay the technical groundwork for future iterations, including offline/edge processing support, expanded multilingual capabilities, and interoperability beyond DHIS2.
  

• Willingness to travel to PATH countries as needed and overlap with GMT and ESA timezones