Responsibilities
- Build and train proprietary models on Oolka's data for repayment-likelihood scoring, negotiation outcome prediction, and credit-risk signals.
- Own the full model lifecycle : data collection, feature engineering, training, validation, deployment, and monitoring.
- Fine-tune LLMs and smaller models for domain-specific tasks, structured extraction from credit reports, and negotiation dialogue quality.
- Build and maintain the evaluation framework that catches model quality regressions before they ship.
- Build feature pipelines from credit bureau, transaction, and repayment data.
- Design and operate model serving : batching, quantisation, versioning, and rollback for models you own.
- Monitor for model drift, degradation, and bias in production, and own the retraining loop.
- Partner with the AI engineering team; they own how models get built and improved; they own how models get served in the live product.
Requirements
- 3+ years building and shipping ML models in production, not just integrating third-party AI APIs.
- Hands-on experience training and fine-tuning models (PyTorch or TensorFlow), classical ML and/or LLM fine-tuning.
- Strong feature engineering and data pipeline experience on structured/tabular data.
- Experience with model-serving frameworks (Triton, TorchServe, TensorFlow Serving) and inference optimisation : batching, quantisation, and distillation.
- Familiarity with MLOps tooling, experiment tracking, model registries, and CI/CD for models (MLflow, Kubeflow, SageMaker, or equivalent).
ML-Specific Expertise
- Built and shipped models predicting real-world outcomes (risk, churn, ranking, or similar); credit, lending, or fraud experience is a strong plus.
- Experience with offline and online model evaluation, held-out test sets, A/B testing, and shadow deployment.
- Understanding of LLM fine-tuning approaches (LoRA/PEFT) and when fine-tuning beats prompting.
- Comfortable with the bias, fairness, and explainability bar that comes with models touching credit decisions.
- Has debugged a model quality regression in production and traced it back to a data or training root cause.
(ref:hirist.tech)