Skip to main content
TALENT247 // SPECIALIST AI EXECUTIVE SEARCH & TECHNICAL RECRUITMENT
ML · Machine Learning

Research Engineer, Multimodal Safety & Post-Training

London, UK · Full-time · Hybrid

AI Specialism

Machine Learning

Workplace

Hybrid

Location

London, UK

Compensation

£125k – £175k

annual package

Role Dossier & Specification

REF: RESEARCH-ENGINEER-

THE CLIENT & MANDATE

Our client is an elite frontier AI research lab in Central London developing flagship multimodal reasoning models deployed to hundreds of millions of users and enterprise developers globally. Their Safety & Alignment Engineering organization ensures that rapidly advancing capabilities in autonomous tool use, multimodal perception, and code synthesis remain robustly aligned, refusal-calibrated, and resistant to adversarial exploitation.

They have engaged Talent247 to recruit a Research Engineer specializing in Multimodal Safety & Post-Training in London. Their team sits directly in the critical path of flagship model launches, building automated red-teaming agents, constitutional preference pipelines, and post-training interventions across text, audio, image, and video modalities.

KEY RESPONSIBILITIES

  • Design and execute post-training alignment interventions—spanning SFT, DPO, RLAIF, and online RL—that enforce nuanced safety specifications without degrading helpfulness or false-refusal calibration.
  • Build autonomous multi-turn adversarial red-teaming agents that probe frontier multimodal models for cross-modal prompt injection, steganographic jailbreaks, cyber-offense uplift, and agentic sandbagging.
  • Train high-precision multimodal reward models and generative safety classifiers capable of reasoning over interleaved visual, audio, and code tool-execution traces.
  • Construct scalable synthetic preference data engines that target rare, high-severity edge cases and ambiguous dual-use scientific queries.
  • Partner with pre-training, post-training, and policy governance leads to define automated launch-readiness safety gates and quantitative risk scorecards.
  • Analyze mechanistic failure modes across training checkpoints to diagnose reward hacking, sycophancy, and deceptive alignment behaviors.

TECHNICAL ENVIRONMENT & OFFERING

Stack: Python, JAX, PyTorch, distributed RLHF/RLAIF frameworks, sandboxed agent evaluation harnesses, large-scale accelerator clusters. Workplace: Hybrid in London, UK. Compensation: £125,000 – £175,000 base salary plus equity/bonus incentives, private medical insurance, pension match, and UK visa sponsorship.

Core Requirements

ESSENTIAL QUALIFICATIONS

  • BS, MS, or PhD in Computer Science, Machine Learning, or a related quantitative discipline with 4+ years of hands-on deep learning and LLM post-training experience.
  • Proven track record implementing SFT, RLHF, RLAIF, DPO, GRPO, or PPO pipelines on large language or vision-language models.
  • Strong systems and ML engineering skills in JAX or PyTorch, including distributed evaluation orchestration and dataset quality diagnostics.
  • Practical experience designing adversarial benchmarks, red-teaming harnesses, or calibrated LLM-as-a-judge evaluation pipelines.
  • Clear analytical mindset capable of translating high-level safety principles into measurable optimization objectives.

PREFERRED / BONUS EXPERIENCE

  • Experience securing multimodal (audio/video/vision) or tool-using computer-use agents against indirect prompt injection and privilege escalation.
  • Publications or open-source contributions in AI alignment, adversarial robustness, representation engineering, or calibration.
  • Familiarity with activation steering, linear probes, or representation-level circuit monitoring during post-training.

Position Telemetry

Location London, UK (Hybrid)
Department Machine Learning
Employment type Full-time
Workplace type Hybrid
Salary range
£125k – £175k annual package