An LLM agent that iteratively self-corrects its outputs

A Python framework for building LLM agents with built-in self-correction loops, enabling more reliable and accurate outputs through iterative refinement.

Multi-objective NAS for efficiency under real deployment constraints

A research project on multi-objective neural architecture search (NAS) that explicitly targets efficiency (parameters/energy) while keeping forecasting accuracy competitive.

Evidence that width should grow faster than depth for transformers

A research project studying architecture-conditioned scaling laws and identifying a critical-depth regime where adding layers can hurt performance.

Research code for my Scientific Reports paper (accepted)

Code, data pipeline and results for a 29-city study of when a zero-shot time-series foundation model can replace a tuned local specialist once compute and energy are priced in.

Data-quality checks before a dataset reaches training

A Python library and CLI that checks schema drift, missingness, duplicates, label leakage and distribution shift, with Markdown, JSON and HTML reports.

Reproducible, publication-ready figure packages

A Python toolkit and CLI that renders research figures at final print width from real results, with versioned journal profiles, captions and a provenance audit.

Distribution-free prediction intervals with coverage guarantees

Wraps any fitted model so it returns prediction intervals or label sets that contain the truth at a chosen rate, with no retraining and no distributional assumption.

Time-to-event analysis on NumPy and SciPy

A lightweight toolkit for right-censored data: Kaplan-Meier, log-rank, Cox models with diagnostics, time-dependent AUC, RMST and competing risks, with a CLI and reports.

Statistical and ML baselines with honest evaluation

A forecasting kit with seasonal-naive, drift, Holt-Winters, SARIMA and Croston baselines, a full accuracy battery including MASE and RMSSE, and a small CLI and pipeline.

Treatment-effect estimators checked against known truth

Estimates treatment effects from observational data with IPW, matching, AIPW, DiD, synthetic control, RD, IV and CATE learners, then scores each against simulated ground truth.

Power, sample size and randomisation for trials and A/B tests

Sample-size and power calculators, CUPED variance reduction, sequential testing, stratified and cluster randomisation, SRM checks and multiple-testing corrections in one small toolkit.

Survey, public-health, and reproducible statistical workflows

Proof page for a Fiverr gig focused on survey data, public-health analytics, and reproducible statistical reporting.

Panel data, time series, and policy-style quantitative analysis

Proof page for a Fiverr gig centered on causal inference, econometrics, forecasting, and robustness-heavy interpretation.

Dashboards that explain what changed, why it matters, and what to do next

Proof page for a Fiverr gig focused on business analytics, KPI reporting, and interpretation-first dashboard work.

One use case, one channel, one integration

Proof page for a Fiverr gig focused on scoped AI workflow automation and internal knowledge assistants.

Legacy Fiverr track: 200+ projects, 4.8+ rating, Level Two Seller

WordPress, Shopify, and Wix development for businesses worldwide. This is the older Fiverr lane that established delivery credibility before the AI/data portfolio.