I build machine learning and LLM systems that are accurate, interpretable, and accountable, from medical image classification to production data pipelines.
Two models you can interact with. One shows its reasoning. The other shows its trade-offs.
Tissue images are classified as benign or malignant by a model that fuses two magnifications. Drag across the slide: at 40X the single-scale model got this malignant sample wrong. At 200X it got it right, which is exactly why the final model uses both.
Drag to compare the slide with the model's attention
SlideGrad-CAM
Drag the dial: how much accuracy does sustainability cost?
λ = 0.0Most recommenders chase the next purchase. Rewear ranks H&M products with LightGBM LambdaRank, then reranks by a sustainability score with a single dial. Drag it to see exactly what each step toward greener recommendations costs in accuracy, with 95% confidence bands.
Where I spend my time, from model design to the pipeline that feeds it.
Explanations built into the model, evaluation weighted toward the errors that cost the most, and limits documented next to every result.
CNNs, Vision Transformers, and multi-scale fusion in PyTorch, with Grad-CAM to show what the network sees.
Learning to rank with LightGBM LambdaRank, multi-source candidates, and evaluation beyond accuracy.
Production LLM pipelines, RAG, Model Context Protocol, and prompt engineering that automate real business workflows.
ETL with quality gates, orchestration, and infrastructure as code using PySpark, Airflow, Postgres, and Terraform.
Time series, regression, and latent class models that answer a decision-maker's question, with the uncertainty shown.
A few recent projects. Each one shows its impact and the responsible AI check behind it.
SARIMA forecasting on 116 years of monsoon data, testing whether forecast-interval breaches can flag floods like 2022.
Impact: SARIMA beat a baseline ARIMA model by 15% on RMSE.
Responsible AI check: A naive historical-average baseline came within 0.03 mm, and the write-up says so.
Scheduled ETL over Denver’s 311 feed: S3 ingestion, PySpark, quality gates, a Postgres star schema, and a live dashboard.
Impact: 69 automated tests and an 80% coverage gate in CI.
Responsible AI check: Quality gates fail the run before bad data ever loads.
A cognitive health app for older adults: five-minute daily brain exercises that start from a 30-day built-in program, then keep going with fresh activities generated by the Claude API.
Impact: A 30-day exercise bank runs fully offline, then the Claude API keeps the program fresh with new daily activities.
Responsible AI check: Claude generates content but never grades the user, and personal data never leaves the device.
I'm glad to talk about interpretable ML, LLM systems, or a dataset that deserves better.