Requirements • 2-3+ years of commercial Python development (FastAPI, async, testing) • 1+ years building LLM applications that ran in production, beyond PoCs • Deep hands-on experience with the Anthropic Claude and/or OpenAI APIs: tool use, structured outputs, prompt caching, and context and cost management • RAG: chunking, embeddings, hybrid search, reranking, and vector databases (pgvector, Qdrant, OpenSearch or similar) • Agentic systems: orchestration, tool calling, multi-step workflows and error handling • Evals and observability for LLM systems (Langfuse, promptfoo or in-house frameworks) • PostgreSQL, Redis, task queues (Celery or similar) and object storage (S3/MinIO) • Docker and Docker Compose, CI/CD (GitHub Actions/GitLab CI) • Experience with AWS (Bedrock, ECS/EKS, Lambda, S3, IAM) • Understanding of data security requirements: PII, access control, audit, and work in closed environments • English at Upper-Intermediate (B2) level or higher Responsibilities • Design client AI solutions and take them to production, from discovery and architecture through deployment and support • Build AI agents, RAG systems, document intelligence (OCR, field extraction, classification) and business process automation • Integrate LLMs with enterprise systems such as Jira/Confluence, Microsoft 365/SharePoint, CRM, ERP/1C, DWH and core banking, using MCP, APIs and connectors • Deliver on-prem and hybrid solutions for regulated clients: local models, PII masking, and work inside the client’s perimeter • Set up quality evaluation: eval sets, accuracy and hallucination metrics, and production monitoring • Take part in technical presales: solution architecture, effort estimates, demos and client technical Q&A • Bring Claude Code and AI-assisted SDLC into client engineering teams, and run hands-on workshops • Review code, set technical standards, and mentor AI Lab engineers Nice to have • Claude Agent SDK, MCP servers, Claude Skills, and Claude Code as part of daily work •…
Джерело
Ми збираємо новини з відкритих джерел і завжди залишаємо посилання на першоджерело.