I’m Sushan Manandhar, a software engineer focused on MLOps. I design the path from a promising experiment to a service that can be checked and operated: versioned inputs, explicit release criteria, useful monitoring, and a way to stop or recover when something goes wrong. My case studies show the decisions and the evidence behind them.
I use FraudShield to show reproducible model delivery, RAGOps to test whether an answer is grounded in evidence, and SignalOps to study risk controls in paper trading. I built ModelGuard to check model artifacts before admission. Each project states what was tested and what remains a deployment design.
I version training inputs, compare a candidate against the current model, and package a typed fraud-risk API. The evaluation uses held-out synthetic data.
MLflowDVCKubernetesTerraformI evaluate retrieval, citations, refusals, safety, and latency before promoting a grounded-answer service. The evaluation set is synthetic.
RAGFastAPIPrometheusI take authorized alerts through deterministic paper-only risk controls, cost-aware simulation, and restart-safe reconciliation. Crypto discovery is a planned research module; no live orders are placed.
Discord BotSQLitePrometheus