Emre Aydolun Pak

Managed a 4-person evaluation pod for Gemini early-access model builds during the June Personal Intelligence update. Owned task execution, reviewer alignment, onboarding, quality checks, and feedback cycles. Acted as the client-facing bridge for personalization behavior, data updates, tool-use workflows, source-selection issues, and model quality feedback. Created Google ecosystem tool-use evaluation examples across Gmail, YouTube, Google Photos, Search context, user preferences, and activity history signals. Evaluated personalization behavior, app-signal routing, context retrieval, grounding, hallucination risks, Turkish localization quality, and incorrect tool-use behavior; refined guidelines, reviewer instructions, rubrics, and localization criteria around user intent and personalization expectations.

Co-led Gemini-based voice-assistant-style workflows focused on tool-use orchestration, instruction following, localization, and reasoning consistency. Designed tool-call workflows, trainer guidelines, calibration sessions, and quality standards. Validated approximately 150 tool-call scenarios including tool selection, execution order, dependency handling, 5-6 step chains, and prompt-grounded completion. Reviewed broken tool-call chains, hallucinations, localization errors, reasoning inconsistencies, and instruction-following failures.

Led a mid-sized global trainer team supporting Yandex SEO algorithm evaluation, search quality analysis, and AI-assisted SEO update workflows. Reviewed trainer-submitted SEO analyses for alignment with Yandex algorithm requirements, search quality guidelines, ranking logic, and structured evaluation standards. Coordinated updates for evolving AI-assisted SEO workflows to keep trainer output consistent with changing ranking logic, keyword interpretation, and relevance criteria. Prepared trainer documentation and E-E-A-T review steps covering author credibility, expertise, trust signals, content ownership, and source reliability. Identified gaps in trainer submissions including weak SEO reasoning, incorrect ranking assumptions, poor keyword interpretation, and algorithm misalignment; improved data quality, documentation clarity, and ranking evaluation consistency.