Melina Thapa — Backend Engineer

$ cat about.md

BSc in Computer Science & Information Technology, Tribhuvan University (2024). Two years building production backend systems for a multi-tenant SaaS ERP — event-driven architecture, CDC pipelines, and entitlement enforcement across tenants. Currently applying that systems background to a research interest in MLOps and trustworthy production ML systems — the project below is where that shift is concrete: the same instincts that keep a multi-tenant backend correct are what built the ML pipeline that watches it.

$ tree credit-anomaly-detection/

Multi-Service Credit & Entitlement Anomaly Detection

Go ingestion service + Python ML pipeline · fully Dockerized (one-command setup)

A multi-tenant SaaS ERP has a quiet failure mode: a single misconfigured entitlement or credit rule can silently affect real revenue across many tenants, long before anyone notices. This project is an end-to-end system for catching those anomalies as they happen — a Go service for high-throughput event ingestion, feeding a Python (scikit-learn, PyOD) modeling pipeline built around Isolation Forest.

The debugging story

The first version of the scoring pipeline used a mean/standard-deviation z-score to flag outliers. It looked reasonable in early testing, but real tenant data is heavy-tailed — a handful of extreme values were dragging the mean and inflating the standard deviation, which meant genuine anomalies were scoring as normal. Switching to a median/MAD (median absolute deviation) estimator — a statistic that isn't dragged around by outliers — fixed detection stability on the skewed distributions this system actually sees in production.

0.753precision

1.0recall

0.859F1

Isolation Forest, held-out evaluation data.

Every flagged anomaly comes with an explanation, not just a score: SHAP-based explainability per anomaly type, exposed through a FastAPI /analyze/{event_id} endpoint that returns a human-readable justification for that specific event.

view source on github

About

Multi-Service Credit & Entitlement Anomaly Detection

The debugging story

Experience

Skills

Contact