PhishGuard: On-Premise AI Email Scam Detection for Government Offices

PhishGuard is a working prototype: a 7-layer email analysis pipeline (sender authentication, domain reputation, psychological-manipulation detection, link analysis, sender history, a Nepal-specific scam pattern database, and a final AI judgment layer) that connects to an office's existing mailbox over standard IMAP no email migration required.

The key design decision, and the one most relevant to a government deployment: the AI component runs entirely on a local, government-controlled server using an open-source model (Llama 3.1 via Ollama) no email content is ever sent to an external cloud API. This directly solves the data-sovereignty problem that rules out commercial cloud-based tools for government use.

When an email is flagged, it's quarantined automatically, and the responsible staff member gets a plain-language explanation (not just a score) plus a real-time notification on their device — including when the AI system itself is uncertain, in which case it's flagged for mandatory human review rather than silently guessing. Staff can also manually report anything suspicious that wasn't automatically caught, feeding back into the system's pattern database over time.

Technical stack: Python/FastAPI backend, Celery task queue, PostgreSQL, React dashboard, fully containerized with Docker for straightforward deployment on office hardware.

This is currently tested on a single mailbox as a proof of concept. What we're seeking through this proposal is ministry sponsorship for a supervised pilot in one real office to validate false-positive/false-negative rates against real government email traffic, expand the Nepal-specific scam pattern database with real cases, and confirm it holds up operationally before any wider rollout.