AI Engineer

About the Role We are building an AI-powered, multi-modal RAN optimisation platform and need a technically sharp junior engineer to help design, train, and deploy language model components at the core of the system. You will work on SLM/LLM selection, fine-tuning, RAG pipeline construction, and production-grade hallucination mitigation. Key Responsibilities Evaluate, benchmark, select, deploy and optimise SLMs and LLMs (e.g., Phi-3, Mistral 7B, Llama 3.x, Qwen 2.5) for telecom-domain tasks including prompt-based optimisation, KPI anomaly explanation, and configuration audit, within on-prem/private cloud environments with GPU acceleration. Design and implement RAG pipelines integrating PM/CM/FM data, drive test logs, and vendor documentation as retrieval corpora; manage chunking, embedding, and vector store selection. Apply LoRA and QLoRA fine-tuning to adapt foundation models on operator-specific network datasets; manage training runs, hyper-parameter sweeps, and evaluation harnesses. Implement and maintain hallucination mitigation strategies: grounded generation, self-consistency checks, retrieval verification, confidence scoring, output guardrails, model drift and prompt failure. Contribute to the model governance pipeline: versioning, shadow-mode evaluation, A/B comparison, and promotion criteria for production deployment. Collaborate with RAN and data engineering teams to ensure model inputs align with real-world PM counter formats, CM schemas, and FM alarm structures. Develop AI agent workflows capable of interacting with telemetry, optimisation engines, RCA workflows, and network automation systems using controlled tool invocation and approval guardrails. Qualifications Requirements: 1–3 years of hands-on ML/NLP engineering experience (internships and research projects count). Strong Python; practical experience with HuggingFace Transformers, PEFT/LoRA, and at least one vector DB (Chroma, Weaviate, pgvector, or similar). Solid understanding of transformer architecture, attention mechanisms, tokenisation, and fine-tuning paradigms (SFT, instruction tuning, RLHF basics). Experience building or productionising RAG systems: document ingestion, chunking strategy, embedding model selection, retrieval evaluation (MRR, NDCG, faithfulness). Familiarity with hallucination failure modes and at least one mitigation approach in production (citation grounding, chain-of-thought, self-RAG, or ROME/MEMIT-style factual correction). Comfortable with experiment tracking (MLflow, W&B) and reproducible training workflows. Familiarity with containerized AI workloads using Docker and Kubernetes. Understanding of GPU scheduling, model serving frameworks (vLLM, Triton, TGI, Ollama, or similar). Awareness of responsible AI practices, model governance, prompt security, and data privacy considerations in enterprise environments. Good to Have: Exposure to structured, semi-structured, and time-series data formats; familiarity with data lakehouse architectures and data pipeline concepts is an added advantage. Experience with quantisation (GPTQ, AWQ, bitsandbytes) for edge or on-premises inference. Published work, GitHub contributions