LLM Developer – AI Safety Guardrails

About the role

Talent Match Africa is recruiting an LLM Developer on behalf of its client. The role owns the model layer that decides whether an AI agent's next action is safe: intent classification on tool calls, prompt-injection and data-exfiltration detection, and risk scoring, all running inline within a latency budget measured in milliseconds. It is applied LLM engineering under real production constraints, with rigorous evaluation and predictable latency as important as accuracy. The position is full-time and on-site in the Nairobi office.

Key responsibilities

• Design and ship the inline guardrail and intent-classification pipeline.
• Build red-team and regression evaluation harnesses for adversarial prompts.
• Fine-tune and distil small models to meet inline latency targets.
• Build retrieval systems over policy, schema and entitlement data.
• Instrument model decisions so that every verdict is auditable and explainable.
• Collaborate with the gateway team on running inference within the enforcement path.
• Balance accuracy, latency, batching and throughput trade-offs in production.

Required profile

• Fluent written and spoken English.
• Ability to present a model you trained yourself, with before and after accuracy and latency results.
• Rigour around evaluation: you build the harness before you trust the number.
• Willingness to work non-traditional local office hours aligned with the client's team, and to collaborate with engineering teams across time zones.
• Detail-oriented, evidence-driven, honest, ethical, transparent and team-oriented.

Required skills

• Strong Python and production codebase experience.
• Transformer fine-tuning, including LoRA / PEFT, distillation and quantisation.
• Model serving with vLLM, TensorRT-LLM or equivalent.
• Frontier model APIs and tool-use / MCP integration patterns.
• Advantageous: security-adjacent ML or adversarial robustness experience, real-time classification or risk-scoring experience, retrieval systems over structured or policy data, and published research or open-source contributions in NLP or ML.