About the Role
Avenga is an international engineering firm helping businesses operate with AI at the core. With 6,000+ experts worldwide, we combine engineering expertise with AI-native thinking to turn ambitious ideas into real-world impact across industries, technologies, and markets. We are actively seeking a professional to strengthen our team in Bulgaria as an AI Software Engineer. You will build and deploy production-grade LLM and agentic AI solutions, working with modern AI frameworks, model APIs, RAG, and cloud infrastructure while helping shape secure and scalable AI solutions for enterprise clients.
Responsibilities
•Integrate frontier model APIs and serve open-weight models on GPU infrastructure, with routing, fallbacks and cost control.
•Build guardrails: input and output validation, PII handling, prompt-injection defences, structured outputs and permissions for tool and agent actions.
•Implement agent and tool protocols (MCP, A2A), and build reference agents and RAG patterns on the platform.
•Benchmark and select models on quality, latency and cost.
•Work with the QA & Evaluation Engineer on evaluation datasets and metrics, and treat evaluation results as the definition of done.
•Work spec-first: write the specification and acceptance criteria, implement with AI coding agents, then review and verify what they produce.
•Keep up with models, protocols and frameworks, and advise the team on what to adopt and what to ignore.
•Deploy the platform with enterprise clients, and contribute to technical proposals.
Requirements
•5+ years of professional software engineering with strong Python; C#/.NET is a plus. Clean, tested, reviewed code is your default.
•Hands-on production experience with LLM applications: model APIs, tool calling, structured outputs, RAG, and at least one agent framework (for example LangGraph, Semantic Kernel, Microsoft Agent Framework or OpenAI Agents SDK).
•Experience running open-weight models, for example with vLLM, TGI or Ollama.
•A practical understanding of how LLM systems fail and how to design against it: hallucination, prompt injection, tool misuse, loops, cost blow-ups.
•Spec-driven development as your normal way of working: specifications, interface contracts and acceptance criteria written before code; daily use of AI coding agents (for example GitHub Copilot, Claude Code or Cursor, with tools such as GitHub Spec Kit or Kiro); rigorous review and testing of generated code.
•Cloud and container literacy: you can deploy what you build.
•Fluent English: you can explain an AI design and its limits to a client.
Nice to Have
•Evaluation tooling: Langfuse, LangSmith, Promptfoo, Ragas or similar.
•Fine-tuning, distillation or quantisation of models.
•Building MCP servers or A2A agents.
•AI security practice, for example the OWASP Top 10 for LLM applications, or familiarity with the EU AI Act.
•Regulated industries, consulting or client-facing delivery.
