Darshini Rajendra

Content Risk Analyst with close to 2 years of experience at Amazon Alexa AI Responsible AI, spanning AI data annotation and quality auditing, LLM judge output validation, adversarial red teaming, customer feedback triage, and incident escalation support. Skilled at applying calibration standards and human-in-the-loop review to evaluate model outputs against defined policy taxonomies, identify systemic failure patterns, and drive process improvements that strengthen safety guardrails. Recognized with Amazon's Core Excellence Award (Q2'25) and Best Debutant Award for identifying violation patterns missed by existing detection logic.

Content Risk Associate — Alexa Sensitive Content & Intelligence (ASCI) at Amazon (2024-09 – Present)

AI Data Annotation, Quality Auditing & Model Improvement

• Conduct structured red-team audits and human-in-the-loop review of Alexa and Metis conversational AI responses, validating outputs against defined content policy taxonomies using rule-level definitions rather than surface-level judgment calls.
• Participate in calibration exercises to maintain reviewer consistency, applying calibration examples and drafting audit findings on policy violation (PV) determinations.
• Flag hallucinations, false deflections, and biased outputs during regression testing on reprocessed datasets under the BEMR (Bad Experience & Risk Management) framework.
• Execute adversarial testing across Alexa conversational AI and image generation/image-to-image editing models, designing multi-turn test cases — an unambiguously clean first turn isolating the exact prompt or edit that triggers a violation in the second turn.
• Identified 8 distinct logic-based policy violation patterns, each tied to a unique evasion or bypass mechanism, directly informing updates to sensitive-content safeguards and reducing downstream model risk.
• Tested Alexa Kids Mode for under-13 safety compliance and validated AI behavior across i18n/multilingual locales spanning North America and Europe, ensuring culturally aligned (not literal-translation) policy interpretation.
• Ran fairness/bias evaluations for demographic neutrality across the Echo Show (5/8/10/15) and Echo Dot device lines.
• Classify and triage customer feedback records against defined SOPs, identifying recurring issues, escalation-worthy trends, and edge cases requiring human judgment.
• Monitor for policy violations and emerging issues, drafting escalation summaries and root-cause hypotheses, and coordinating with cross-functional teams to drive timely resolution.
• Built and deployed Alexa Red Team Risk Analyzer, an agentic AI tool automated on Amazon Q that converts analyst intake requests into structured red-team scenarios — Risk Scenario Description, Policy Category, Adversarial Technique, Failure Mode, Endpoint, Severity, and Testing Use Case — standardizing scenario design and cutting manual drafting time across the team.
• Extended the tool to automate consolidation of red-teaming data across the 8-member team, replacing a manual process of merging individual Excel trackers with a single automated consolidated output.
• Awarded Core Excellence Award (Q2'25) and Best Debutant Award for sustained accuracy in policy calibration and identification of novel violation patterns missed by existing detection logic.

Frontend Developer & UI/UX Design Intern at Kamya Hastakala Arts (2024-02 – 2024-05)

• Built a responsive e-commerce site (React.js, HTML/CSS) with secure payment integration and a Figma-designed UI, informed by usability testing.

Bachelor of Engineering in Information Science & Engineering – New Horizon College of Engineering (2024)