Zendesk

Machine Learning Engineering Intern · Oct 2025 – Present · Berlin

Zendesk is a global customer experience platform used by over 100,000 companies. I joined the Berlin AI team as an ML engineering intern, working on the agentic AI product that automates customer support at scale.

I designed and built an AI agent security framework that became the pre-deployment security gate for all of Zendesk's AI agents ahead of their public GA release. The framework covers over 1,500 attack scenarios — prompt injection, jailbreaks, policy violations, and adversarial inputs — and is now standardized across all agent releases.

I also designed a LLM-as-a-Judge online scorer that reaches 92% agreement with human raters and deployed it in Braintrust to continuously monitor agent quality in production. This replaced manual QA sampling with a fully automated, always-on quality signal.

Security Framework

4 attack categories, pre-deployment security gate for all AI agent releases. Standardized across the entire Zendesk agent platform.

1,500+ SCENARIOS

PROMPT INJECTION → JAILBREAK → POLICY VIOLATION → ADVERSARIAL INPUT → DETECT → CLASSIFY → BLOCK → REPORT → PROMPT INJECTION → JAILBREAK → POLICY VIOLATION → ADVERSARIAL INPUT → DETECT → CLASSIFY → BLOCK → REPORT →

PHASE 01CRITICAL

Prompt Injection

Manipulating agent instructions

PHASE 02HIGH

Jailbreaks

Bypassing safety guardrails

PHASE 03MEDIUM

Policy Violations

Breaking business rules

PHASE 04HIGH

Adversarial Inputs

Edge-case exploitation

QUALITY SCORERDEPLOYED IN BRAINTRUST

92%agreement with human raters

Replaced manual QA sampling with always-on quality monitoring. LLM-as-a-Judge evaluates every agent response in production.

Fully automated, continuous quality signal