Research Engineer, Sequential Intelligence

The company

Embreier builds Kalyr, the consequence record for enterprise AI. For every recurring decision in an operation, human or AI, Kalyr seals the forecast before action, records what was permitted and done, and settles the outcome after it in the company’s own economics. Platforms run the work; Kalyr proves what it earned, and learns from it. We are building intelligence that learns from reality wherever consequential work is done.

The role

You build the models that predict how a real process responds to each move a team could make, and the evaluation that tells us, plainly, whether they did. You turn messy operating history into clean decision state, honest baselines and calibrated predictions, and you ship them.

Observation, inference and confidence stay distinct from the first line of code. You decide what to build and, as often, what to leave out. The questions are hard and specific: “What information actually existed before this decision?” “Are these two changeovers truly comparable?” “How do we know this model beats the method already installed?” The systems you shape become the foundation everything else stands on.

What you will do

• Predict the next state. Sequence and tabular models that forecast how a process moves under each candidate action, with calibrated ranges.
• Seal before the world answers. Record every prediction before its outcome exists, in a form anyone can audit later.
• Evaluations that tell the truth. Chronological, site-held-out and regime-held-out splits, proper scores, calibration, and abstention when evidence is thin.
• Hunt leakage. Find hindsight leakage before it finds you, and make the harness catch it every time.
• Beat the installed method, honestly. Score every model against the strongest alternative, including the method the plant already runs.
• Run the learning loop. Decide which scored episodes teach the next model, with human approval, clean update and rollback, and which of one site’s learning holds at the next.
• Code others can enter. Build it so another excellent engineer can enter the codebase without reverse-engineering it from your head.

What you bring

• Systems that shipped. Production work on sparse, sequential or longitudinal data, owned end to end, in fluent, typed and tested Python.
• Modelling depth. Probabilistic modelling, time series, state-space or representation learning, and strong tabular baselines.
• Evaluation rigour. Walk-forward validation with purging and embargo, proper scoring rules, calibrated uncertainty. You know how a benchmark flatters, and how to stop it.
• Disciplined curiosity. Real interest in causal and counterfactual questions, held with care.

What you will work with

Python, PyTorch and the scientific stack; gradient boosting and tabular foundation models as baselines; conformal and Bayesian uncertainty; PostgreSQL and a replay harness built for exact chronology.

Strong additions: off-policy or uplift evaluation, sequential decision-making, world-model evaluation, industrial or energy data.

This is for you if

• You want your models judged by reality, settled in money, beyond any leaderboard.
• Long, intense stretches of building give you energy, and you stay precise and grounded deep into them.
• You ship. Your work runs in production and other people build on it.
• A result that disappoints you is information, and you say so first.

To apply

The application asks two questions in place of a cover letter. What is the hardest system you have personally built, what broke, and what would you change? And: Which assumption in Kalyr’s learning loop would you try to break first, and how would you test it? A strong answer is a test you could run, not an opinion. We read the answers before the CV.

Who we hire

Original thinkers. People who learn a rule well enough to follow it, then bend it to build something better. You contribute in work others can build on, and you measure yourself by it. You are glad to be known by what you create and by what you give. Proud of the work, humble in the room. The shared result comes first. Accountability here means being the first to say the true thing that needs saying, for the good of the mission we share.

Life at Embreier

Embreier is a Swiss company based in Lausanne. Engineering, research and people who know operations from the inside build in one team, across languages and disciplines. High standards and appreciation live side by side here: the work is demanding, and the people doing it are seen.

Need an adjustment to apply? Write to careers@embreier.com and tell us how we can help.