AI Scientist

What Makes This Role Different

We’re not training just another LLM. Our goal is to give models the ability to act autonomously, running code, fixing systems, crunching data so humans can focus on strategy and creativity. If turning state-of-the-art research into something that does real work excites you, read on.

Your mission

You’ll research, prototype, and ship features that move our Agentic LAM from proof-of-concept to production. Below is the menu of topics you may touch. You don’t need expertise in all of them—bring depth in one or two, plus the curiosity to learn the rest.

1. Foundations of ML & AI

• Transformers & Attention — scaling large language models (LLMs)
• Sparse / Modular LLMs — Mixture-of-Experts (routing, gating, model-parallelism)
• Retrieval-Augmented Generation (RAG) — search-enhanced generation pipelines
• Supervised Fine-Tuning & Distillation — DeepSeek-style recipes, LoRA/QLoRA
• Reinforcement Learning for Alignment — RLHF/RLAIF, RLVR, GRPO
• Emergent Abilities & Scaling Laws — data-compute trade-offs, power-law trends
• Safety & Guardrails — constitutional prompting, adversarial red-teaming

2. Mathematical Bedrock

• Linear algebra — low-rank factorisations, efficient projections
• Probability & information theory — KL control, entropy regularisation
• Optimisation — Adam/Adafactor vs. SGD, second-order & constrained RL methods

3. Model Lifecycle & Tooling

• Data curation → training → evaluation (BLEU, Rouge, TruthfulQA, bespoke suites)
• Behavioural tests — toxicity, drift, hallucination, jailbreak resilience
• Experiment tracking & reproducibility — W&B, MLflow, GitOps

4. Data Design & Reasoning-Driven Distillation

• Chain-of-Thought & Self-Consistency prompting
• Dataset distillation / condensation for lean retraining
• Instruction & rationale generation from teacher models
• Curriculum & slice-based sampling (easy→hard, semantic buckets)
• Active learning loops with human or automated feedback

5. Code & Runtime Stack

• Programming languages — Python, C/C++, Go
• Training engines — PyTorch, JAX/Flax, PEFT/LoRA
• Distributed compute — DDP, FSDP, Ray, Kubernetes
• Inference — INT8/4-bit quantisation, ONNX, Triton, vLLM, streaming APIs
• Async services — FastAPI, gRPC, task queues

(We supply the GPUs; you decide how to light them up.)

What We Look For

• Solid Python and familiarity with at least mutiple topics above.
• Appetite to dive into papers or repos, run quick experiments, and share learnings.
• Comfort with the chaos & freedom of a startup: shifting priorities, small teams, big impact.
• No formal degrees required—we hire for skill, not certificates.

Why You Might Thrive Here

• Ownership from day 1 — your code will steer real agents, not sit in a backlog.
• Mentorship loop — ex-Google/Amazon engineers and published researchers on call.
• Open scope to grow — conferences, open-source, new research lines.
• Equity upside — help shape the company, share the reward.

(Startup life isn’t for everyone: pace is fast, guardrails are few, but the learning curve is vertical.)

Working Model & Benefits — Fully Negotiable

• Hybrid in Lisbon + multi-week remote blocks
• Company-paid travel to global AI events
• Competitive salary + equity
• Subsidised apartment if you relocate
• Health insurance, Modern hardware, learning budget
• NDA only — no non-compete

If this role excites you, but you are worried that you don't fit all the requirements, please send your application anyway. We would love to get in touch!