Research Engineer

Artificial Societies (societies.ai) helps Fortune 500 organisations understand how real-world audiences think, feel, and respond — without the cost, delay, or limitations of traditional research. We build large-scale simulated populations grounded in real-world data, enabling clients to test messages, strategies, and concepts. Our clients love the ability to receive insights backed by millions of responses in hours rather than months, and to access previously inaccessible audiences (e.g. investors, CEOs, opinion leaders).

We're backed by Y Combinator, Point72 Ventures, and Kindred Capital, plus investors from DeepMind and Sequoia Scout, and work with leading organisations across technology, financial services, communications, and public affairs. We're a lean and highly effective team. We move fast, hold a high bar, and trust each other to own outcomes end-to-end. The culture is collaborative but low-ego: ideas win on merit, not seniority.

The Role

We're looking for Research Engineers to work on our core simulation engine — the system that builds AI digital twin personas, assembles them into societies, and simulates their attitudes and opinions across situations. This is not a pure research role, and it's not a pure engineering role. We need people who can move fluidly between designing rigorous evaluations, training and fine-tuning models on cloud compute, prototyping novel architectures, and shipping production code. You'll own the full research-to-production lifecycle: from formulating a hypothesis about how to better capture human nuance in a language model persona, through to deploying and monitoring the system that makes it real.

Our simulation engine poses hard problems. How do you construct an AI persona that coherently captures the beliefs, contradictions, and reasoning patterns of a real human? How do you compose thousands of such personas into a society that reflects real-world diversity and group dynamics? How do you simulate opinion distributions using language models in a way that’s statistically faithful and practically useful? How do you design reward functions that steer model behaviour toward human-faithful outputs rather than superficially plausible ones? These are the questions you’ll work on every day.

We maintain a tight research-to-product pipeline. Our simulations power real high-stakes decisions for enterprise clients — the research has to be rigorous, and it has to ship. We’re looking for people who find it gratifying to see their work pushed to its absolute limits.

What You’ll Do

• Build and evaluate our simulation engine — design, implement, and improve the systems that construct AI personas from real-world behavioural data. You’ll work on how personas reason, how they express opinions, and how they respond to stimuli in ways that are faithful to their real-world counterparts.
• Train and fine-tune foundational models — work with the latest GPUs to train, fine-tune, and adapt language models for persona simulation. This includes designing reward functions and training objectives that steer model outputs toward human-faithful behaviour rather than generic or superficially plausible responses.
• Architect societies — develop the methods by which individual personas are composed into populations that capture meaningful human diversity, group dynamics, and emergent collective behaviour. This includes how opinion distributions arise, how sub-groups interact, and how aggregate signals remain statistically grounded.
• Write the right evals — design and build evaluation frameworks that go beyond standard benchmarks. You’ll develop rigorous statistical methods to prove the fidelity of our simulations — measuring how well our synthetic populations mirror real-world opinion distributions, demographic patterns, and response dynamics.
• Advance the state of the art — stay at the frontier of relevant research across deep learning, language model alignment, synthetic data, and computational social science. Reproduce, critique, and improve upon academic work. Translate theoretical breakthroughs into production-ready improvements.
• Ship production code — this is not a notebook-only role. You’ll write clean, well-structured Python that runs in production, reviewed and merged alongside the engineering team using Git and code review. Your research has to work at scale.

What We’re Looking For

• Strong experience with deep learning and NLP. You understand transformer architectures deeply — not just how to call an API, but how attention mechanisms work, how to fine-tune effectively, how to design training objectives and reward functions, and how model behaviour changes under different training regimes. You’ve trained or fine-tuned models on cloud compute (e.g. AWS, GCP) and are comfortable working with distributed training infrastructure.
• Extremely high proficiency in Python. You write clean, performant, well-structured code. You’ve shipped code to a production codebase — not just prototypes or notebooks.
• Evaluation mindset. You have strong instincts for what to measure and how to measure it. You can design evals that reveal whether a system is actually working, not just whether it looks like it’s working.
• Solid mathematical and statistical foundations. You’re comfortable with probability, statistics, information theory, and the quantitative reasoning needed to assess whether a simulation is faithful to reality. You can reason formally about distributions, divergence measures, and calibration.
• Creative problem-solver. You come up with novel architectural or methodological approaches — not just incremental improvements. You can look at a problem from first principles and propose something new.
• Production engineering habits. You’ve worked in a shared codebase with other engineers. You use Git fluently. You review code. You understand that research code that can’t be maintained is research that doesn’t ship.
• Ownership mentality. You don’t wait to be told what to do. You see what needs to happen and you make it happen. You care about the outcome, not the process.

Nice to Have

• Experience or genuine interest in computational social science, social data science, or behavioural modelling — understanding how opinions form, propagate, and shift in populations.
• Published research or meaningful open-source contributions in relevant areas.

Logistics

• Location: London, in-office. We build better together, and this role is no exception.
• Equity: 0.03% – 0.20%.

Why Join Now

This is a rare opportunity to work on one of the most intellectually ambitious problems in applied AI: building faithful models of human attitudes and opinions — and making those models useful for the highest-stakes decisions in the world. You’ll work directly with the founders, shape the technical direction of the company, and build systems that no one else has built. If you want to do research that ships and engineering that matters, this is the role.