PhD quantitative researcher. Economics, statistics, mathematical modelling, and — in public code — machine learning, SQL, RAG evaluation, and PyTorch.
The economics work is not a previous identity. Identification, information sets, and independent checks are the same habit whether the object is an IV, a nested-CV score, a SQL feature, or a RAG trace. Job history is quantitative research, not an ML-engineer title.
Economics Expert / PhD Quantitative Research · Econometrics, causal inference, forecasting, optimisation, AI evaluation of quantitative answers.
Machine Learning Expert / Computational STEM · Python, model validation, ground-truth engineering, SQL, GenAI/RAG, PyTorch, reproducible serving.
Samples in the Python repos are documented simulated DGPs or closed forms unless a file says otherwise. Recovering a simulated parameter is not an empirical finding. Nothing here is a commercial deployment.
econometrics-causal-inference-lab — DiD, IV, RD, matching, and panel methods on documented DGPs. A coefficient that prints is not an identified treatment effect. CASE_STUDY.md
computational-ml-stem-problem-forge — twelve problems, each with a reference solver and two further checks that are not copies of that solver. FLAGSHIP_CASE_STUDY.md
statistical-reasoning-validation — Type I / Type II, coverage, and p-value misuse under known DGPs. statistical_error_catalogue.md
machine-learning-model-selection-lab — invalid workflows kept next to the matching design: full-frame scaling, group leakage, inner best_score_ treated as generalisation. CASE_STUDY_MODEL_SELECTION_FAILURE.md
ai-response-evaluation-benchmarks — fluent answers that fail on the target, the information set, or the interpretation. One author coded the YAML. FLAGSHIP_REVIEW_CASE.md
genai-rag-evaluation-lab — gold can sit at rank 1 while the extractive answer abstains. Retrieval metrics are not answer quality. FLAGSHIP_RAG_FAILURE_ANALYSIS.md
- sql-ml-feature-engineering-lab — a join without
txn_ts <= cutoffadmits a planted99999.0without selecting the label. - pytorch-deep-learning-lab — hand derivatives, finite differences, and autograd; they disagree on purpose at a ReLU kink.
- mlops-reproducible-serving-lab — HTTP 200 with the wrong probability after a column swap.
- time-series-forecasting-lab — walk-forward skill, not in-sample fit.
- optimization-decision-models — solver success is not KKT.
- quantitative-finance-models — identities under stated assumptions, not a trading book.
Lean 4 work (compilation is not faithfulness): lean4-optimization-economics, lean4-formalization-review, lean4-formalization-faithfulness, lean4-proof-engineering, lean4-analysis-formalization, lean4-automation-debugging, lean4-mean-value-theorems.
Python, R, SQL, NumPy, Pandas, SciPy, statsmodels, scikit-learn, PyTorch, Lean 4/mathlib.