Data Scientist

Do you have strong mathematical intuition, excellent programming skills, and a passion for solving complex problems? Are you interested in adapting large language models using proprietary datasets and turning research into production applications?

At Mobilewalla, you will work with large-scale data and a multidisciplinary team to develop models that solve challenging consumer intelligence problems. We are particularly interested in candidates with strong foundations in statistics, mathematics, and rigorous experimentation.

About Mobilewalla:

Mobilewalla is a consumer intelligence company headquartered in the United States, with a global presence across Europe, APAC, and ANZ. We develop computational and statistical techniques that transform large-scale data into actionable consumer insights, helping marketers make informed decisions and execute data-driven campaigns.

Our work spans efficient data processing, statistical modelling, machine learning, and consumer intelligence. We combine data science and engineering expertise to solve challenging problems and deliver solutions at scale.

About the Role:

You will develop and adapt machine learning models, with a particular emphasis on fine-tuning large language models using Mobilewalla’s proprietary datasets. Your responsibilities will span training data preparation, model selection, fine-tuning, statistical evaluation, and production deployment.

This role requires more than familiarity with LLM tools. We seek candidates who can reason mathematically about model behaviour, design sound experiments, diagnose failures, and demonstrate measurable improvements in model quality and business outcomes.

Key Responsibilities: LLM Fine-Tuning and Domain Adaptation

• Adapt pretrained language models to domain-specific tasks using proprietary structured and unstructured datasets.
• Design and run supervised fine-tuning experiments using full fine-tuning or parameter-efficient methods such as LoRA and QLoRA, as appropriate.
• Assess when continued pretraining, instruction tuning, preference optimisation, or retrieval-augmented generation offers the strongest approach.
• Build training datasets through careful curation, labelling, deduplication, quality filtering, and representative sampling.
• Develop data splits that prevent leakage across related entities, time periods, and near-duplicate examples.
• Investigate how training data composition, model size, hyperparameters, and optimisation choices affect performance and generalisation.
• Evaluate adapted models against base models and simpler alternatives to establish the value of fine-tuning.
• Diagnose overfitting, catastrophic forgetting, hallucinations, and performance degradation across tasks and data segments.

Statistical Modelling and Experimentation:

• Apply probability, statistical inference, linear algebra, and optimization to formulate problems and guide model development.
• Design controlled experiments with appropriate baselines, uncertainty estimates, and statistical tests.
• Develop evaluation strategies that address sampling bias, class imbalance, noisy labels, and distribution shifts.
• Build predictive models for applications such as segmentation, propensity prediction, forecasting, and anomaly detection.
• Use error analysis and ablation studies to distinguish meaningful improvements from experimental noise.
• Translate model performance into measurable business outcomes and communicate assumptions and limitations clearly.

Production Delivery and Collaboration:

• Collaborate with data and software engineers to build reproducible training, evaluation, and inference pipelines.
• Optimise model deployment for quality, latency, throughput, GPU memory requirements, and cost.
• Monitor production performance and establish criteria for model updates and retraining.
• Work with product and business stakeholders to define requirements, prioritise experiments, and deliver useful capabilities.
• Evaluate relevant research and incorporate methods that demonstrate practical value.

Required Qualifications and Skills:

• At least 2 years of professional experience in data science, machine learning, or statistical modelling.
• Bachelor’s, master’s, or doctoral degree in Statistics, Mathematics, Computer Science, Physics, Engineering, Operations Research, or a related quantitative discipline, or equivalent demonstrated expertise.
• Strong foundations in probability, statistical inference, linear algebra, calculus, and numerical optimisation, with the ability to apply these concepts to practical modelling problems.
• Strong Python programming skills and proficiency in SQL.
• Hands-on experience training or fine-tuning transformer-based language models on custom or domain-specific datasets.
• Understanding of transformer architectures, training objectives, tokenisation, gradient-based optimisation, and regularisation.
• Experience with PyTorch and the Hugging Face ecosystem, including Transformers, Datasets, and parameter-efficient fine-tuning tools.
• Ability to prepare high-quality training data and design evaluations that measure generalisation and prevent data leakage.
• Experience with statistical analysis and traditional machine learning using libraries such as SciPy, statsmodels, scikit-learn, or XGBoost.
• Familiarity with GPU-based training, experiment tracking, and reproducible development practices.
• Ability to explain modelling choices, interpret experimental results, and solve ambiguous problems through first-principles reasoning.
• Strong collaboration and communication skills.

Preferred Qualifications:

Preference will be given to candidates with substantial mathematical or statistical training and demonstrated depth in experimental reasoning. Evidence may include advanced coursework, research, publications, or technically rigorous industry projects.

Additional preferred experience includes:

• Fine-tuning LLMs on proprietary enterprise datasets and deploying the resulting models in production.
• Continued pretraining, preference optimization methods such as DPO, or reinforcement learning approaches for model adaptation.
• Distributed and memory-efficient training, including mixed precision, gradient checkpointing, and multi-GPU training.
• Training data selection, sample efficiency, learning curves, and systematic analysis of data quality.
• Assessing model calibration, uncertainty, robustness, and performance under distribution shifts.
• Privacy-preserving machine learning and evaluation of training data memorization.
• Large-scale data processing with Apache Spark and cloud infrastructure such as AWS.
• Consumer intelligence, advertising technology, audience modelling, causal inference, or recommendation systems.

What Success Looks Like

• Fine-tuned models demonstrate reproducible improvements over base models and relevant alternatives on representative, held-out evaluations.
• Proprietary datasets are converted into high-quality, well-documented training and evaluation assets.
• Model improvements are supported by sound statistical evidence and translate into measurable business value.
• Production models meet agreed standards for reliability, privacy, latency, and cost.
• Research findings and technical decisions are clearly documented and shared across teams.

Why Join Mobilewalla?

• Work on domain-specific LLM development using large-scale proprietary datasets.
• Apply mathematical and statistical thinking to challenging, real-world problems.
• Own the model development lifecycle—from data preparation and experimentation to production impact.
• Collaborate with a multidisciplinary team across global locations.
• Help shape how advanced AI supports consumer intelligence and data products.

Skills

Statistical Inference, Retrieval-Augmented Generation (RAG), Training Set, Training Design, Optimisation Techniques, QLoRA, Domain Adaptation Techniques, Scientific Expertise, Fine Tuning, Data Domain, Hugging Face, Tokenization, Computational Modelling, Language Models, Dealing with confidential data, Programming