MSc research · Evidence published

Telecom operations anomaly detection

Engineering research into safe handling of operational telemetry: anonymisation, layered data, tests and run controls. Supports the analytics career as engineering depth — not as the primary market lane.

Layered pipeline from bronze tokenised ingest through silver and gold anomaly scores to a review view — Layered engineering flow — bronze / silver / gold with privacy and run controls.

Business context

Operational networks generate high-volume telemetry. Useful detection work depends on data that can be handled safely and reproducibly.

The problem

Without clear extraction controls, anonymisation and testable pipelines, anomaly work is hard to defend and risky to share.

My role

Designed and implemented Python workflows with layered data, tests and explicit run controls as MSc engineering research.

Approach

Safety-conscious data handling and anonymisation patterns (site tokens only).
Layered datasets: bronze ingest → silver clean/DQ → gold anomaly scores.
Automated tests and documented run controls, including a dry-run gate.
Publish labelled stills from a synthetic dataset — not operator environments.

Architecture / model

See the layered flow above. Evidence below is pipeline behaviour (volumes, scores, tests) — not a management dashboard.

Business logic (in the stills)

Layer volumes retained across bronze / silver / gold
Mean anomaly score by day with a review threshold
Silver DQ review flags
Automated test and run-control board

Controls

No raw operator IDs — invented site tokens only
Schema, privacy, DQ and model tests recorded as Pass / Warn
Dry-run gate and “production credentials absent” check
Every visual watermarked as synthetic / public-safe

Result

Published synthetic layered evidence. Public repository link remains withheld until disclosure and governance are cleared.

Bronze silver and gold layer row volumes — Layer volumes

Mean anomaly score by day with review threshold — Anomaly score timeline

Run controls and automated tests board — Tests and run controls

Source tables (CSV):

What I am learning

How to keep detection work reviewable: privacy first, then layers, then scores — with tests that fail closed when safety gates are not met.

Public-disclosure note

No production credentials, access methods or identifiable operator environments are published here. Figures and site tokens are invented for portfolio evidence.