AI Engineer

Location: Dubai, United Arab Emirates (Micropolis Robotics)About the RoleWe’re building an AI Assistant that serves as a digital concierge, helpdesk, and knowledge hub inside our mobile application. Users will:Ask questions about our services and their own data (documents and tables)Perform actions by invoking our application services through the assistant (tool use/MCP/API integrations)Converse via both text and real-time voiceThe AI system runs using open-source models and tooling, with on-premise deployment as a first-class option.Key ResponsibilitiesDesign, build, and operate a production server-side AI assistant that:Answers questions grounded in user-scoped data (docs, tables) with citations where applicablePerforms actions by calling internal/external APIs securely on the user’s behalfSupports low-latency, real-time voice chat (streaming STT/TTS + incremental LLM responses)Implement the tool/agent layer:Structured tool calling (JSON-schema based) to integrate business servicesModel Context Protocol (MCP) servers/clients where appropriate for tool discovery and executionArchitect retrieval-augmented generation (RAG):Ingestion for documents and tables, parsing, chunking, embeddings, metadata, and indexingHybrid retrieval (sparse+dense), query rewriting, and answer attributionDeliver performant, cost-efficient inference on open-source models:Model selection/routing; context management; caching/batching; streaming token deliveryGPU utilization and serving via vLLM/TGI/llama.cpp/Ollama or similarBuild resilient APIs and real-time integrations:WebSockets/WebRTC/gRPC for streaming voice; REST/GraphQL for control and orchestrationProductionize and operate on server/on-prem:Containerize with Docker; automate CI/CD; implement logs/metrics/traces (OpenTelemetry)Evals, A/B tests, safety/guardrails, and human-in-the-loop feedbackRequired Skills and ExperienceProficiency in PythonOpen-source LLMs and serving:Experience with models like Llama, Mistral/Mixtral, Qwen, etc.Serving stacks: vLLM, Text Generation Inference (TGI), llama.cpp, OllamaPrompt engineering, routing, context/window managementRAG and data systems:Vector DBs (like FAISS, Qdrant, Weaviate, Milvus, pgvector) and hybrid searchDocument/table ingestion and normalization; schema/metadata designReal-time voice:STT/TTS (e.g., Whisper/faster-whisper, Vosk, Coqui TTS, Piper) with streaming pipelinesLow-latency streaming via WebSockets/WebRTC/gRPCTooling for actions:Structured tool/function calling, API design/integration, and service orchestrationFamiliarity with Model Context Protocol (MCP) concepts and usageDeployment and operations (on-prem first):Docker, Linux, networking, and secure service deploymentGPU stacks (CUDA/drivers/containers) and performance tuningExcellent communication, documentation, and cross-functional collaborationPreferred SkillsAgents/frameworks: LangChain, LlamaIndex, Semantic Kernel, or custom tool routersAdvanced retrieval: multi-vector stores, RRF/hybrid search, query planning, re-rankingSQL generation and safe execution over tabular data; row-level security; schema mappingDocument processing: OCR, table extraction, CSV/Parquet pipelinesServing/perf: Triton, quantization (GGUF/GGML), LoRA/QLoRA with PEFT, KV-cache optimizationsEvals and observability: Ragas/DeepEval, Langfuse/PromptLayer, OpenTelemetryNice to Have (Not Required)Public cloud experience (AWS/Azure/GCP) for hybrid or future deployments#J-18808-Ljbffr