Hayk M.

Production Data & Decision Systems Engineer | ML-backed systems

Most data and machine learning projects do not fail because the models are wrong. They fail because real decisions are built on systems that were never designed for production reality. Data distributions change. Assumptions stop holding. Feedback loops appear once decisions are automated. What looks correct during analysis can quietly become wrong when it starts affecting real users, money, or operations. My work focuses on that gap between analysis and real-world execution. I design and implement production decision systems, not demos, notebooks, or isolated models. These are systems where predictions directly drive actions such as approvals, rejections, prioritization, allocations, or automated workflows. In this context, correctness, transparency, and operational control matter more than theoretical performance. My background combines applied machine learning, data engineering, and direct ownership of real decision pipelines. I have worked on systems used in environments where mistakes are costly, including credit decisioning and automated trading. I approach every project assuming that data will be imperfect and conditions will change over time. Models are components of a system, not the system itself. I work system-first. That means clearly defining decision logic, data contracts, monitoring signals, and conditions for retraining or rollback. A system should surface degradation early, before it causes damage, not after. What I do not do is equally important. I do not sell AI magic, guaranteed outcomes, or opaque black-box solutions. I do not optimize offline metrics without considering how the system behaves once deployed. If a system cannot be monitored, understood, and controlled in production, I do not consider it complete. Clients usually reach out to me when decisions matter, scale is real, and failure is expensive. If you need a system that behaves predictably under pressure and you want to clearly understand both its capabilities and its limits, we are likely a good fit.