Senior Machine Learning Engineer (Generative AI)

Kilwa; Chicago, Northern (IL, KY); Hybrid; USD 120,000 - 180,000; Full time

Job summary

Kilwa Technologies is seeking a senior ML engineer to own the model layer behind ISI, METI and Zawadi, designing and productionising a hybrid ensemble that scores market inputs.

You will work on the retrieval/generation layer and the explainability surface, ensuring every decision is traceable and backed by backtesting and drift-detection tests.

Qualifications

• Proven production ML at depth with ownership over deployed models.
• Fluency across gradient boosting, sequence models and transformer NLP.
• Preference for published, interrogable methodology over raw accuracy.

Responsibilities

• Design, train and productionise the hybrid model stack behind ISI and METI signals.
• Build the retrieval and generation layer for Zawadi so every answer traces back to a scored input.
• Own the explainability surface — SHAP attribution, model cards, and the invalidation conditions we publish.
• Set the evaluation bar: backtesting, drift detection, and robustness tests that appear in report appendices.

Skills

Production ML ownership

Gradient boosting

Sequence models

Transformer NLP

Explainability expertise

Tools

SHAP

Model cards

Job description

Own the model layer behind ISI, METI and Zawadi — the hybrid ensemble that turns multilingual, data-scarce market inputs into scores institutions will act on.

Location Chicago or remote (US)

Type Full-time

Team Engineering

What you'll do

• Design, train and productionise the hybrid model stack behind the ISI and METI signals.
• Build the retrieval and generation layer for Zawadi so every answer traces back to a scored input.
• Own the explainability surface — SHAP attribution, model cards, and the invalidation conditions we publish.
• Set the evaluation bar: backtesting, drift detection, and the robustness tests that appear in report appendices.

What you bring

• Production ML at depth, with real ownership of a model that people made decisions on.
• Fluency across gradient boosting, sequence models and transformer-based NLP — and judgement about when each is the wrong tool.
• A bias toward published, interrogable methodology over unexplained accuracy.