The posting
Job Title: Senior AI Product Engineer
Location: Belgrade, Novi Sad - Serbia or Lisboa/Porto/Braga, Portugal
Working Model: Hybrid (flexible, depending on candidate location)
Reports to: sa.global Labs Leaderhip
Department: sa.global Labs – Core Team (team working on AI development within sa.global)
Seniority Level: Senior
empower AI Platform - Senior AI Product Engineer, Agentic Systems
About the Role
If you've ever shipped an agent that worked great in the demo and fell apart in production, this is the role that fixes that — for real clients, at real scale.
sa.global Labs is looking for a Senior AI Product Engineer to own the reasoning core of empower, our industry-specific agentic AI platform: the agent orchestration, prompt architecture, and evaluation systems that turn agentic behavior into automation clients can actually trust. You'll build on LangGraph, DSPy, and MLflow, reasoning over a Neo4j-backed knowledge graph and hybrid retrieval layer, with real latitude to decide when agentic complexity is warranted and when it isn't.
If you're a Senior AI Platform Engineer whose interest keeps pulling toward the agent layer, this is where that curiosity turns into ownership.
About empower AI Platform
empower is a proactive agentic intelligence layer. It's built around 3 pillars, and this role touches all of them, with a focus on the applied AI/agent layer:
- Corporate Knowledge (Ontology): the structured semantic model of a client's domain that gives agents grounded, domain-correct context.
- Data Hub: the layer that unifies, cleans, and serves enterprise data into the platform.
- Multi-Agent Skills: the orchestration layer where reasoning, tool use, and multi-step automation happens.
Platform Stack
empower is built on a specific, opinionated stack. Prior hands-on experience with these, not just the general category, is what we're screening for:
- LangGraph: agent orchestration and control flow.
- DSPy: declarative prompt programming and optimization.
- LlamaIndex: retrieval and RAG pipeline framework.
- PostgreSQL with pgvector and pg_search extensions: our production vector store and hybrid full text/semantic search layer.
- Neo4j: the graph database underlying the Business Knowledge Graph.
- MLflow: experiment tracking and lifecycle management for models and prompt/agent evaluation runs.
- Prefect: orchestration and scheduling of data and processing pipelines.
This role builds the reasoning and agentic core of empower, the backend of the platform that decides, plans, and acts, and that has to be evaluated and trusted rather than just shipped and hoped for.
Core responsibilities
- Design and implement multi-agent workflows and orchestration graphs using frameworks such as LangGraph.
- Apply DSPy or comparable declarative prompt-programming approaches so that prompting is systematic, testable, and optimizable, not hand-tuned strings that rot with every model update.
- Build and maintain evaluation harnesses — tracked in MLflow — that measure accuracy, faithfulness/hallucination rate, latency, and instruction-following across model and prompt versions.
- Architect agent memory, tool-calling, guardrails, and human-in-the-loop controls for production agent workflows serving real client decisions.
- Prototype retrieval architectures on top of the Neo4j-backed Business Knowledge Graph (GraphRAG, hybrid search with LlamaIndex and pgvector/pg_search) in partnership with Level I engineers.
- Push technical judgment calls on when an agentic pattern is warranted versus when a simpler deterministic path is the right call, this role is expected to have and defend that opinion.
Required technical experience:
- 5+ years of backend software development; Python as the primary language for AI-adjacent work.
- Proven track record of designing and shipping production APIs and distributed services, not just prototypes.
- Hands-on experience integrating LLM/AI APIs (Anthropic, OpenAI, or open-weight models via vLLM/Ollama) into real applications. This is application integration, not model research.
- Working knowledge of retrieval architectures built on PostgreSQL with pgvector and pg_search (our production stack), chunking, hybrid vector/full-text search.
- Familiarity with MCP (Model Context Protocol) for exposing tools and data to AI agents.
- Comfort with cloud-native deployment (Azure preferred, given our Azure DevOps stack; AWS/GCP experience transfers fine), containerization, and CI/CD.
- Daily, hands-on use of Claude Code or a comparable AI coding agent (Cursor, Copilot, Codex) as a working method for scaffolding, refactoring, test generation, and codebase navigation, not just occasional use — using it to prototype and stress-test agent architectures, generate evaluation scaffolding, and build custom skills/tooling around the AI coding agent itself.
- Deep, hands-on production experience with LangGraph or a comparable agent-orchestration framework.
- Working knowledge of LlamaIndex (or a comparable RAG framework) for building retrieval pipelines over enterprise data.
- Experience using MLflow (or a comparable tool) to track experiments and evaluation runs across model/prompt versions.
- Solid grounding in LLM evaluation methodology and a genuine point of view on how to measure a non-deterministic system.
- Experience designing agent architectures end-to-end: planning strategies, memory, tool use, guardrails, and failure recovery.
Nice to have
- Hands-on experience with Neo4j. Comfortable modeling and querying graph-shaped data (Cypher), and reasoning about how ontology structure affects retrieval and agent grounding.
- Hands-on experience with Prefect (or a comparable orchestrator like Airflow) for building and scheduling data and processing pipelines.
- Real working experience with DSPy or a similar framework for prompt optimization and compilation, not just having read about it.
- Fine-tuning or preference-optimization experience for domain-specific tasks.
- Open-source contributions to agent or LLM-orchestration frameworks.
- Experience translating whitepaper-level agentic-AI concepts into shipped products (adapting external frameworks or research into a platform's own terminology and architecture).
Soft Skills & Working Competencies
Technical range alone won't succeed on this team. empower is built by a small, AI-native team working across genuinely different domains, often ahead of settled best practice, and increasingly through AI coding agents rather than only with them. The competencies below are assessed in interviews, not treated as filler. Each comes with a working definition so there's no ambiguity about what's expected.
Agency
The capacity to identify what needs to happen and act on it without waiting to be told, spotting a gap, defining the goal, and mobilizing the tools (including AI agents) and people needed to close it, while owning the outcome. This is distinct from raw autonomy: autonomy is being able to work unsupervised; agency adds the initiative to decide what is worth doing next. In this role, it looks like: spotting that a retrieval pipeline is producing weak grounding, proposing the fix, and driving it to done, including deciding when an AI coding agent can execute the plan and when it can't.
Systemic Thinking
The ability to reason about a component in terms of its effect on the whole system, not just its local correctness, understanding how a change in one part of empower (a retrieval change, an agent policy, an ontology edit) propagates through data flows, other agents, and end-client outcomes. It includes seeing feedback loops and second-order effects, not just the immediate diff.
Structured Communication for AI-Directed Work
The ability to write clear, well-scoped instructions, specs, and delegation, to teammates and AI coding agents alike, precisely enough that the recipient (human or model) doesn't have to guess. Includes decomposing large problems into well-bounded tasks and knowing what to delegate versus what to do by hand.
Ownership & Quality Discipline
Treating agent behavior, evaluation results, and prompts as your responsibility regardless of whether a human or an AI agent wrote the first draft, building the feedback loops (evaluation harnesses, tests, review) that keep quality high when a growing share of the work is AI-generated.
Comfort with Ambiguity
The ability to make sound technical progress in areas where the tooling, frameworks, or client requirements are still settling, common in agentic AI and in translating enterprise ontology work across service-centric domains.
For more information, visit www.saglobal.com.
