Key Responsibilities

• Advanced Prompt Engineering: Designing complex, dynamic prompt templates with conditional logic and efficiently reusing information and context within prompts to maximize generation quality and reasoning. - Structured Outputs & Schemas: Implementing various response schemes (JSON mode, function calling, Zod/JSON schemas) to ensure AI outputs are predictable and ready for seamless integration into application logic. - Prompt Engineering & Evaluations: Building robust evaluation pipelines and using Langfuse to collect feedback and score the quality of responses in real time. - Tracing & Debugging: Performing deep debugging of complex LLM chains using Langfuse traces to identify bottlenecks and optimize for cost, latency, and context window usage. - AI A/B Testing: Running systematic experiments across different models via OpenRouter (e.g., comparing Claude 3.5 Sonnet vs. GPT-4o) and analyzing results based on quantitative metrics. - Data-Driven Decisions: Making deployment decisions for new prompts or models strictly based on quantitative benchmarks and trace data, rather than intuition. - Output Scoring & Analysis: Developing scoring systems to analyze the “Problem → Solution” chain and identify root causes of hallucinations or logic errors using Langfuse analytics. - Model Performance & Fine-Tuning: Regularly re-evaluating model performance as new architectures emerge and performing fine-tuning when necessary to meet specific domain requirements. # Qualifications

• Node.js & Next.js: Deep knowledge of the stack to build reliable services and handle complex LLM-generated data. - Dynamic Prompting Skills: Proven experience in building prompts where content is highly dependent on input variables and context injection. - OpenRouter Experience: Experience working with unified APIs, managing rate limits, and selecting the most cost-effective models for specific tasks. - Langfuse (or similar): Understanding of LLM observability principles — setting up tracing, creating test datasets, and integrating scoring systems. - Evaluation Methodology: Experience with frameworks like RAGAS or building custom “LLM-as-a-judge” systems. - Analytical Mindset: Ability to transform raw generation logs into actionable business metrics and technical insights. - Iterative Mindset: Focus on continuous product improvement through constant feedback loops. - Fluency in Russian and/or Ukrainian. #