VLM Research Engineer London ·

You'll take open-weight vision-language models and post-train them for specific intelligence tasks on images and video, owning the work from data to production.

About Sidney: Sidney is an AI intelligence platform that runs investigations the way an experienced analyst does. It plans, collects across open, deep and dark sources, grades every source, reasons over the evidence and states plainly what it could not establish, with a human analyst at the decisions that matter. Sidney is live, with a growing base of customer and trial users. Analysts rely on it in their day-to-day work for deep research, person investigations, corporate due diligence and image and video analysis, and what they need shapes what we build next.

What you'll be involved in:

• Own specialist models. Choose the base model and the training recipe for each task, and make the call on when a model is ready to ship.
• Post-train it properly. Run supervised fine-tuning, preference optimisation and reinforcement learning with verifiable rewards, and know which one the problem actually needs.
• Make evaluation the bar. Build the held-out benchmarks and baselines that decide whether a model ships, and keep tracking them after it does, so regressions are caught before an analyst sees them.
• Ship and serve it. Train and deploy on AWS, wire the model into Sidney's investigation pipelines, and keep latency and cost-per-query where the product needs them.
• Watch it in production. Trace and monitor model behaviour with LangSmith or similar tools, and turn what analysts flag into the next round of training data.

What we're looking for:

The person:

• You've post-trained a model that shipped. Not prompting, not a notebook. You've taken a model through fine-tuning to real users, and you can say what moved the number and why.
• Rigorous about evaluation. Held-out sets, baselines, leakage, and knowing when a number is real. You'd rather ship nothing than ship a model that isn't measurably better.
• Research taste, engineering habits. You read papers and know which ideas are worth trying here, and you write code that others can run, review and maintain.
• Strong communicator. You explain clearly, in conversation and in writing, what a result does and doesn't show, to engineers and non-engineers alike.

Technical skills:

• Post-training. Hands-on supervised fine-tuning of vision-language or multimodal models, parameter-efficient methods such as LoRA and QLoRA, and preference or reinforcement-learning methods such as DPO or GRPO.
• Stack. Strong Python and fluency in at least one deep learning framework, such as PyTorch or JAX.
• AWS. Training and serving models on AWS: GPU instances, SageMaker or Bedrock, and S3-based data pipelines.
• AI monitoring. Experience with LangSmith or similar tools, such as Langfuse, Weights & Biases or Arize, for tracing, evaluating and monitoring models in production.

Nice to have:

• OSINT, remote-sensing or other image-intelligence work.
• Text recognition in low-quality images and video, including non-Latin scripts.
• Serving open-weight models efficiently in production, with vLLM or similar.
• Distributed training across multiple GPUs.

What you'll get: We're a small, performance-driven team at one of the fast-growing startups in AI for intelligence work.

• Performance-driven, low ego. The best argument wins, not the loudest voice or the longest title. We give direct feedback, expect it back, and hold each other to a high bar.
• Ownership from day one. You'll own a real part of Sidney, not a ticket queue. What you decide ships to customers in weeks, not quarters.
• Judged on impact. We care about what you ship and what it changes, not hours or visibility. Strong work is noticed quickly and rewarded with more scope and responsibility.
• Growth that's real. Sidney is growing fast, and so are the roles inside it. Early hires shape how engineering works here and take on more as we scale.
• Work that matters. Your work lands in front of analysts making high-stakes calls, on a product that's already live.

How to apply: The application form takes about twenty minutes. You'll need:

• Your CV, as a PDF or Word document.
• Links to your LinkedIn, GitHub or something you've shipped, if you have them.
• Short written answers to 6 questions about this role, in your own words.

No cover letter. A founder reads every application, and we reply to all of them.