Abi Aryan is the founder of Abide AI, a machine-learning research engineer, and the author of O’Reilly’s LLMOps: Managing Large Language Models in Production. She builds and studies intelligent systems that must remain reliable, interpretable, and affordable beyond controlled benchmarks.
Aryan studied mathematics and computer science, investigated image recognition with Hopfield networks, and moved from business intelligence and data analytics into machine-learning research. As a visiting research scholar at UCLA, she worked under Judea Pearl on intelligent agents. Her subsequent engineering work encompassed recommendation systems, automated audio and video labeling, speech synthesis, forecasting, computer vision, and natural-language processing.
In 2019, Aryan coauthored Arena, a platform for studying how multiple reinforcement-learning agents cooperate, compete, and adapt. She later investigated production machine learning’s operational challenges, including data infrastructure, monitoring, debugging, and coordination between teams. Her research on language-model generalization and cost-efficient deployment extended those concerns into generative AI; her 2025 book formalized LLMOps around model selection, evaluation, deployment, security, observability, and infrastructure.
- Evaluation beyond static benchmarks. Generative models produce multiple plausible answers, making conventional accuracy insufficient. Aryan’s work on language-model evaluation examines how application-specific metrics, automated checks, model-assisted assessment, and human judgment can establish practical trust.
- Domain adaptation under real deployment constraints. Her AI Engineer Summit talk distinguishes adapting models to an entire domain from tuning them for individual tasks. She compares prompting, retrieval, instruction tuning, and parameter-efficient fine-tuning, emphasizing that LoRA and four-bit QLoRA depend on hardware compatibility, software dependencies, memory budgets, and clean training data.
- Agent evaluation in changing environments. Aryan coauthored AbideGym, which tests reinforcement-learning agents against changing conditions, and research on causal reflection exploring how language-model agents reason about actions and outcomes.
- Hardware-aware inference optimization. Her public account of moving into GPU engineering connects her work at Abide AI with hardware, compilers, distributed systems, on-premises models, and the practical economics of deploying agentic software.
