Denys Linkov

Denys Linkov is a machine-learning leader at Wisedocs and a University of Toronto lecturer whose work focuses on making AI reliable in medical-document processing and software engineering. He has helped develop Medical Long Context Reasoning, built machine-learning teams at Voiceflow and Wisedocs, and evaluated whether coding agents can safely refactor complex production systems.

From conversational AI to healthcare

Linkov came to machine learning through cloud engineering and became Voiceflow’s first machine-learning hire in 2021. Over approximately three and a half years, he built and led its machine-learning team, developing a custom MLOps platform, fine-tuned encoder models, retrieval-augmented generation services, and production microservices. He subsequently became an adviser.

At Wisedocs, he applied that experience to medical-claims files that can exceed 10,000 pages. As commercial models and open-source infrastructure improved, his priorities shifted toward combining encoder and decoder models, strengthening healthcare expertise, and making automated decisions accountable to human reviewers. His work on managed document-verification agents describes systems that pair automated analysis with human quality auditors.

Linkov also teaches machine learning at the University of Toronto and develops LinkedIn Learning courses on building AI applications. His CSC490 course connects students with industry practitioners, whom he has publicly thanked for contributing to classroom instruction.

• Medical Long Context Reasoning: Linkov and Wisedocs developed MLCR, a benchmark examining whether models can answer medical-claims questions accurately as irrelevant contextual material accumulates. The framework spans more than 250 questions across six difficulty levels; public synthetic cases and an evaluation harness are complemented by private, harder questions designed to limit contamination. Linkov announced its release with Artificial Analysis.

• High-confidence coding-agent evaluation: Linkov argues that agents should be judged at success thresholds closer to 90% or 99%, not the 50% rates often emphasized in task-horizon comparisons. A failed hour-long autonomous run wastes compute and human attention; useful deployment requires clear specifications, verification, and human accountability.

• Production refactoring with measurable business value: While rebuilding a legacy medical-claims pipeline, his team evaluated five orchestration projects against 17 criteria, prototyped with Temporal, and consolidated more than ten repositories into a monorepo. The overhaul lowered processing costs and runtime, accommodated larger files, broadened developer participation, and reduced some feature-delivery timelines from months to less than a week.

• Domain-first AI team design: Linkov organizes teams around actual delivery bottlenecks, favoring adaptable generalists early and specialists when concrete technical constraints demand them. He expects engineers to understand customers, domain experts to help define evaluations, and teams to retain the operational context necessary to verify increasingly autonomous systems.