Randall Hunt is the chief technology officer of Caylent, where he guides enterprise AI and cloud systems from promising prototypes into production. His approach combines deployment discipline, developer experience, and hard scrutiny of whether an application serves its users affordably.
Hunt worked at NASA and 10gen, the company that became MongoDB, before leading a continuous integration and delivery team at SpaceX. He subsequently spent five years at Amazon Web Services in software engineering and technical evangelism; his AWS writing covered services including Lambda, Aurora, Kinesis, and SageMaker. At Meta, he led developer advocacy for PyTorch, the open-source machine-learning framework.
He joined Caylent in January 2022 as vice president of cloud strategy and solutions. By 2024, he was vice president of cloud strategy and innovation, coauthoring practical guidance on evaluating language models, before becoming CTO.
What production AI actually demands
- Evaluation before orchestration. Hunt begins with clear inputs, expected outputs, and tests that expose failures as models evolve. A single successful example can seed an evaluation set; simple pass-fail checks often outperform elaborate scoring. His case for disciplined prompt engineering, coauthored with Guille Ojeda, favors improving instructions and context before introducing elaborate agent architectures. For BrainBox AI, Caylent helped develop ARIA, an assistant supporting building operations, HVAC management, and decarbonization.
- Multimodal search grounded in retrieval. For Nature Footage, Hunt described indexing wildlife videos with Amazon Nova, pooled frame embeddings, captions, and Elasticsearch. Sports footage yielded another practical shortcut: detect crowd reactions with FFmpeg and annotate court boundaries to help vision models recognize plays. He favors Postgres with pgvector and OpenSearch because production search also requires metadata filters, appropriate indexes, and manageable infrastructure costs.
- Interfaces designed around actual users. Nurses working in noisy hospitals preferred text to voice; technicians with limited connectivity needed individual document pages instead of enormous downloads. For CloudZero, Caylent developed generative interfaces for cloud costs, dynamically producing React components and visualizations tailored to users’ questions.
- Unit economics as architecture. Hunt prioritizes prompt caching, lean context, batching, model selection, and conventional computation for deterministic arithmetic. He introduced Bedrock Battleground, Caylent’s open-source model-comparison and prompt-cost tool, while crediting Jon Spaeth as its primary contributor.
His recent work extends those principles into Caylent Accelerate: specialized agents for database migration, software delivery, and cloud operations, combined with deterministic scripts, human oversight, and deployment within customers’ AWS environments.
