Job Description

What we do​

​

msg global solutions leverages cutting-edge technologies, including AI, to solve business challenges. We partner with industry leaders like SAP and AWS to deliver enterprise solutions to Fortune 500 companies, transforming data into valuable insights.

Our solutions reshape industries such as finance, insurance, and manufacturing, specializing in:

• Advanced analytics and AI technologies
• Industry-specific technologies and IoT integration
• Data-driven operational excellence

By partnering with us, organizations can transform their business models, uncover new revenue streams, and achieve operational excellence. As part of the msg group, we combine specialized agility with global IT resources to shape the future of business.

What you will do​

You build the things that actually do the work. Starting from a business process and the people who run it, you decide how that process breaks down into steps a model can handle, what context each step needs, which tools it may call, when a human has to be asked, and how anyone can tell whether the result is correct.

This is applied engineering rather than research. You will not be training foundation models. You will be making general-purpose ones reliable enough that a business is willing to let them touch real work, which is a harder problem than it sounds and a different skill entirely.

Accuracy is the number you own. When something works 80% of the time and has to work 98% of the time, closing that gap is the job.

• Turn business processes into working agents: decompose the process, define the steps, the tools, the handoffs and the points where a human has to decide.
• Design and build retrieval pipelines: chunking and indexing strategy, embedding model selection, hybrid dense and keyword search, cross-encoder reranking, and grounding that keeps answers tied to real source data.
• Treat prompt and context engineering as an engineering discipline, with versioning, tests and measurement rather than tweaking by feel.
• Build the evaluation layer: golden datasets defined together with business experts, automated scoring, model-graded evals where they are appropriate, and regression suites that run on every change.
• Instrument production systems so quality is observed rather than assumed: traces, token cost, latency, failure rates, and drift as data and models change underneath you.
• Own quality end to end. Diagnose whether a failure came from retrieval, reasoning, tool design or the underlying data, and fix the actual cause.
• Choose models and manage the trade-off between quality, cost and latency, and revisit those choices as the available models change.
• Work directly with business experts to extract the domain logic that exists only in their heads and turn it into something testable.
• Take agents into production with the platform and infrastructure team, and stay accountable for how they behave once they are live.

What we are looking for​

​

Requirements​

• 5+ years building software, including 2+ years on LLM-based systems that reached production and real users.
• Strong Python and the habits of a working engineer: version control, tests, code review, incremental delivery. Comfortable with async work, FastAPI and Pydantic or their equivalents.
• Hands-on with at least one agent orchestration framework (LangGraph, LlamaIndex Workflows, OpenAI Agents SDK, Temporal or comparable) and a clear view of what belongs in the framework and what belongs in your own code.
• Experience serving models through a managed platform (AWS Bedrock, Azure AI Foundry, Google Vertex AI) or directly against provider APIs, including streaming, structured output, prompt caching and rate-limit handling.
• Production vector search with at least one of pgvector, Qdrant, Pinecone, Weaviate, Milvus or Azure AI Search, and an opinion on the trade-offs between them.
• Deep practical retrieval experience, and a clear account of why naive RAG underperforms and what you did about it: chunking, metadata filtering, query rewriting, hybrid search, reranking.
• Fluency with tool and function calling, structured output and multi-step agent design, including how each of these fails.
• Hands-on experience with MCP (Model Context Protocol): building and consuming servers, designing tool interfaces that a model can use reliably, and handling authentication, permissions and error behaviour across them.
• Real evaluation discipline. You can describe the eval harness you built for a previous system, what it measured, and what the numbers were before and after your work.
• Comfort with ambiguity: able to spend an hour with a business expert and come away with a workable specification.
• The judgement to know when a language model is the wrong tool, and the willingness to say so when a rules engine or a plain query would do the job better.
• Fluent English.

Technology you will work with

Language and services

Python, FastAPI, Pydantic, async workers

Orchestration

LangGraph, or comparable frameworks such as LlamaIndex Workflows, OpenAI Agents SDK or Temporal

Model access

Frontier and open-weight models via AWS Bedrock, Azure AI Foundry or Google Vertex AI, and direct provider APIs

Tool integration

MCP servers and clients, provider-native function calling, structured output; A2A where agents span platforms or teams

Retrieval

pgvector, Qdrant, Pinecone or Azure AI Search; hybrid dense and keyword search; cross-encoder reranking

Evaluation

Langfuse, LangSmith or Phoenix; Ragas or purpose-built scoring; regression suites running in CI

Delivery

Docker, Kubernetes, Terraform, GitHub Actions, OpenTelemetry

Nice to have​

• Fine-tuning or LoRA adapter training, together with a clear view of when it beats better prompting and retrieval.
• Self-hosting open-weight models: vLLM or similar serving stacks, quantisation, GPU cost management.
• Multi-agent orchestration patterns and their failure modes.
• Agent-to-agent interoperability with A2A: Agent Cards, the task lifecycle, and delegation between agents that sit on different platforms or belong to different teams.
• Document understanding at depth: OCR, layout-aware extraction, tables and semi-structured content.
• Designing the interaction between a person and an agent: how corrections, rejections and escalations are handled.
• Domain exposure in [relevant industries].
• Open-source contributions, technical writing or conference talks.

What we offer​

• A place where individuals are equally valued and where diversity and cultural differences are cherished.
• A global team of highly respected SAP and industry experts where you can make a difference.
• Competitive salaries (55.000 – 65.000 Eur) and a broad range of benefits, some of which are highlighted below.