A leading internet provider is looking for a Senior AI Engineer to join their team in Middelburg, Mpumalanga. 

The Mission:
We are building a sovereign, company-wide AI fabric. This system ingests every document, ticket, transaction, and data point across our portfolio (Veralogix, Bioniq, Innovata, Treadstone, Aiguille) into a unified retrieval, property-graph, and agentic layer. It provides natural-language, grounded, and traceable insights in real-time.

You will own the end-to-end architecture as part of a small, high-impact engineering team, focusing on on-prem-first deployment and seamless integration with our spatio-temporal platform.

Responsibilities: 

1. The Retrieval Spine

• Design and implement a chunking strategy for heterogeneous content: PDFs, scanned docs, transcripts, tables, code, and ERP records.

• Engineer hybrid search (BM25 + dense + late-interaction), reranking, and query routing.

• Architect and operate the vector database infrastructure (pgvector + Qdrant or equivalent).

2. The Corporate Knowledge Graph

• Design and implement the ontology for the corporate knowledge graph (Memgraph, Neo4j, or KuzuDB).

• Execute entity resolution across systems like Odoo, Aleph, Hoover, Elasticsearch, Nextcloud, and telemetry.

• Build GraphRAG patterns that solve complex, multi-hop queries more effectively than naïve vector retrieval.

3. The MCP Server Fleet

• Author and manage a fleet of MCP servers to expose internal systems (Odoo, Hoover, Aleph, Memgraph, Elasticsearch, etc.) to the agent layer.

• Implement robust scoping, authentication, auditing, and rate limiting.

4. Agentic Orchestration

• Build and deploy multi-agent systems using frameworks like LangGraph, Pydantic AI, DSPy, or BAML.

• Implement planner/researcher/executor/verifier patterns with structured outputs, error recovery, and long-running stateful workflows (e.g., using Temporal).

5. Eval, Observability & LLM-Ops

• Build an evaluation harness from scratch with golden sets and RAGAS-style metrics.

• Establish a rigorous “LLM-as-judge” methodology to block deploys on regression.

• Set up observability (Langfuse/Phoenix), cost/latency dashboards, and a version-controlled prompt management system.

Requirements: 

• Production ML/AI: 5+ years of experience, with ≥2 years specifically on LLM-era systems (RAG, agents, fine-tuning).

• Production RAG: Designed and operated a system over heterogeneous content. Can defend decisions on chunking, hybrid retrieval, and reranking from first principles.

• Vector Databases: Deep operational knowledge of pgvector, Qdrant, Weaviate, or LanceDB. Has opinions on HNSW vs. IVF, quantization, hybrid indices, and index rebuild strategies.

• Property Graphs: Production experience with Memgraph, Neo4j, or KuzuDB. Fluent in Cypher, has designed an ontology, and built production GraphRAG retrievers.

• Agentic Tooling: Authored MCP servers or has equivalent depth with production tool-use design (OpenAI function calling, Bedrock agents).

• Agentic Frameworks: Production experience with LangGraph, Pydantic AI, DSPy, or BAML. Has shipped multi-agent systems with real tool calls and recovery logic.

• Evaluation Discipline: Has built an eval harness from scratch and can defend the metrics and regression rates of their prompt changes.

• Information Extraction: Proficient with NER/RE models and structured-output pipelines (Pydantic/JSON schema/BAML) beyond simple prompting.

• Python Mastery: Operating at a staff-engineer level. Can read research papers and implement them.

• Model Fine-Tuning: Experience with LoRA/QLoRA or full fine-tuning of open-weight models.

• Advanced Retrieval: Production experience with ColBERT, SPLADE, or similar late-interaction/ sparse-dense hybrid systems.

• Novel RAG Architectures: Shipped systems like GraphRAG (Microsoft), Contextual Retrieval (Anthropic), or HippoRAG.

• Entity Resolution at Scale: Experience with Splink, Zingg, dedupe, or similar libraries.

• Local Model Serving: Skilled with vLLM, TensorRT-LLM, SGLang, or llama.cpp.

• Spatio-Temporal RAG: Experience with natural-language querying over PostGIS + TimescaleDB.

• OSINT Workflows: Integrated with tools like Maltego, SpiderFoot, Aleph, or Hoover.

• Agentic Coding: Regular user of Claude Code or similar tools as a default working mode.

• Open Source: Contributions to BAML, DSPy, LangGraph, or comparable libraries.

We will contact you telephonically in 3 days should you be suitable for this vacancy. If you are not suitable, we will put your CV on file and contact you regarding any future vacancies that arise.