Skills
What I’m best at
My core strength is statistical modelling and uncertainty-aware inference (particularly Bayesian hierarchical methods) applied to problems where data are noisy, sparse, biased, or structured across space and time. That foundation shapes everything: careful assumptions, principled model selection, rigorous validation, and outputs that are honest about what they do and don’t know.
I’ve extended that same rigour into AI engineering: building LLM-powered applications, RAG systems with proper evaluation, agentic workflows, and fine-tuned models. The instincts transfer directly: if it can’t be evaluated, it can’t be trusted, whether that’s a Bayesian forecast or an LLM pipeline.
What I deliver
- End-to-end AI systems: RAG pipelines with retrieval evaluation (MRR, nDCG, LLM-as-judge), agentic workflows with tool calling and stateful backends, and fine-tuned models (Frontier supervised fine-tuning and QLoRA for open-source LLMs) — built with structured outputs and production habits from the start.
- Full-cycle ML and modelling: problem framing, feature engineering, model selection, rigorous validation, and reproducible delivery — with particular depth in Bayesian hierarchical inference, spatiotemporal forecasting, and uncertainty-aware decision support.
- Data pipelines and analytics: clean, testable, version-controlled pipelines from raw data to decision-ready outputs, with SQL and Python as the core tools.
- Evaluation and communication: leakage checks, calibration, error slicing, robustness testing — and results communicated with explicit assumptions, trade-offs, and clear recommendations.
Core stack
AI / LLM OpenAI · Anthropic · Google · HuggingFace · LangChain · LangGraph · ChromaDB · RAG pipelines · QLoRA fine-tuning (Llama 3.2, 4-bit NF4/LoRA) · Frontier fine-tuning · tool calling · agent orchestration · multimodal inputs (text, image, audio) · Modal serverless deployment · Weights & Biases · Gradio · Streamlit
ML / Data Python (pandas, NumPy, scikit-learn, XGBoost, TensorFlow/Keras, PyTorch) · SQL (PostgreSQL) · R (tidyverse) · Git/GitHub
Methods Bayesian & hierarchical modelling · spatiotemporal forecasting · supervised/unsupervised ML · NLP and embeddings · ranking and recommendation · A/B testing · causal inference · model evaluation and interpretability
What I bring from research
- First-principles thinking: breaking open-ended problems into well-formed questions with the right method for the data-generating process.
- Measurement and study design: sampling protocols, data quality standards, bias and missingness at the source — not as afterthoughts.
- Delivery under real constraints: scoping work, managing trade-offs, and shipping high-quality outputs on time.
- Transparent communication: clear, auditable narratives with explicit assumptions, evidence, and limitations — the standard I held in peer-reviewed work, applied to every project.
- Cross-functional collaboration: working across disciplines, mentoring others, and building alignment through clear writing and structured proposals.
Technical depth
Bayesian, hierarchical & spatiotemporal modelling
- Generalised linear models (GLMs) and generalised additive models (GAMs) for nonlinear effects
- Threshold and segmented regression for decision-point inference
- Hierarchical and mixed-effects modelling; partial pooling
- Bayesian inference with uncertainty quantification and propagation; priors as explicit assumptions
- Spatiotemporal modelling: structured dependence, forecasting logic, species distribution modelling (SDMs)
- Observation vs process modelling: detection–abundance separation; N-mixture models
- Integrated Population Models (IPMs)
LLMs, prompt engineering & agents
- Prompting for structured outputs; reliability patterns (prompt scaffolds, self-checks, evaluation loops)
- Tool calling and retrieval patterns; schema/contract design for model outputs
- Agent workflows: planning/acting loops, orchestration, retries, human-in-the-loop checkpoints
- LangGraph concepts: state, control flow, tracing/debugging
- Multi-model experimentation: Frontier APIs (OpenAI, Anthropic, Google) and open-source models via HuggingFace
- Multimodal inputs: text, image, and audio processing in LLM pipelines
- RAG: vector embeddings, open-source vector datastores, retrieval pipeline design
- Fine-tuning — Frontier: supervised fine-tuning of closed models for domain-specific tasks
- Fine-tuning — Open-source: QLoRA fine-tuning; training open-source models to match or exceed Frontier performance on specific tasks
- LangChain for LLM application orchestration
- Production deployment patterns: end-to-end productionised LLM systems with agentic capabilities
Machine learning (classical)
- Supervised learning: linear/logistic regression, tree-based models, random forests, gradient boosting (incl. XGBoost), SVM, k-NN
- Unsupervised learning: PCA, clustering (K-Means), anomaly detection
- scikit-learn Pipelines; hyperparameter tuning; model comparison and baselines
Deep learning
- Neural networks for classification and regression
- Training fundamentals: loss functions, optimisers, regularisation, monitoring and early stopping
- TensorFlow / Keras / Pytorch: model definition, training, evaluation
- Fine-tuning and parameter-efficient training: QLoRA for open-source LLMs
Evaluation, interpretability & reporting
- Validation design: train/val/test, cross-validation, temporal/blocked splits where appropriate
- Evaluation discipline: leakage checks, calibration awareness, error analysis and slicing, robustness/stress testing
- Interpretability: feature importance, partial dependence, SHAP-style global/local explanations
- Decision-ready reporting: assumptions, limitations, tradeoffs, and clear recommendations
Experimentation & causal inference
- A/B testing fundamentals: hypotheses, metrics, power and effect size
- Causal inference basics: confounding, selection bias, counterfactual framing, limits of identification
- Practical decision-making under uncertainty: interpreting results and communicating tradeoffs
Data engineering & integration
- Data ingestion and transformation: structured files, schema discipline, reliable I/O
- SQL-centric data work: joins across complex relational datasets, analytics transformations
- API integration patterns: extracting, normalising, and joining external data sources
Software engineering & reproducibility
- Git/GitHub workflows: branching, pull requests, code review, merge discipline
- Maintainable codebases: modular architecture, clean interfaces, reusable components, pipeline-style structure
- Quality controls: input validation, assertions, unit tests (pytest patterns), docstrings, type hints where useful
- Reproducibility: environment management (conda/venv), deterministic runs, versioned artefacts, methods-first documentation
NLP, recommenders & text features
- Text preprocessing and inspection: normalisation, tokenisation, feature auditing
- Vectorisation: bag-of-words, TF-IDF, and dense embeddings (SBERT, sentence-transformers); semantic search and similarity-based retrieval
- Recommender foundations: similarity metrics, collaborative filtering, constraint-aware framing
Cloud (AWS)
- Foundational understanding of AWS cloud concepts: shared responsibility, high availability, scalability/elasticity, and core tradeoffs
- Core AWS services: EC2, Lambda, S3, EBS, EFS, RDS, DynamoDB
- Networking & access fundamentals (high level): VPC, subnets, security groups, NACLs, Internet Gateway, NAT Gateway
- Observability & governance (high level): CloudWatch, CloudTrail, AWS Organizations, Cost Explorer, Budgets
Geospatial & remote sensing
- Raster/vector workflows; spatial joins; geoprocessing pipelines
- Spatial feature engineering; landscape/canopy metrics
- Scalable spatial processing
Visualisation & lightweight apps
- Visualisation: matplotlib, seaborn, plotly; ggplot2
- Lightweight apps: Streamlit (Python), Gradio (Python), Shiny (R)
Education & Training
A detailed, certificate-linked list of formal education and courses is maintained here: Education & Training.