Work
Small, well-defined pieces of machine learning and data work, finished.
I take contract work in data analysis, forecasting, and AI systems. I am most useful on the jobs that are genuinely hard rather than merely long: the dataset nobody can get into a usable shape, the model that fits the training set and nothing else, the pipeline that works until the data changes.
What I take on
Data cleaning and analysis
Messy, nonstandard, or partly broken data turned into something you can compute on. Exploratory analysis with an actual conclusion at the end rather than a wall of plots.
Forecasting and time series
Regression, time series analysis, and forecasting, including data with gaps, sensor drift, and structure the standard methods assume away.
Machine learning
Model selection, training, evaluation that means something, and deployment. Classical methods where they win, which is more often than the current mood suggests.
Retrieval and agent systems
RAG over your own documents, multi-agent workflows, hierarchical routing, and automation with n8n. Built so you can run it without me afterward.
Handover
I train the people who will own the work after I leave. This is the part most contractors skip, and it is why a system stops getting used three months later.
Case studies you can poke at
The churn rate is the wrong number
Survival analysis on student retention, a frailty model for why the churn rate falls when nothing about any student changes, and lifetime value recomputed off the survival curve. Sliders move the assumptions and the decision flips.
Who to call is not who will leave
Targeting retention outreach by churn probability performs worse than random. Uplift modelling on a randomised trial, a Qini curve you can read at any depth, and the intervention cost where targeting starts to pay.
Both come out of one repository that works the problem end to end: acquisition and funnel, survival and lifetime value, a discrete-time hazard model, causal uplift, seasonal decomposition and forecasting, hierarchical pooling across centers, drift monitoring, and retrieval evaluation. Eight write-ups, 38 tests, and a command-line tool that validates its input before it predicts.
The data throughout is artificial, for privacy. The analyses are the ones I ran against the real records. Code: github.com/Wassaw-io/tutoring-center-analytics.
One more example
At NASA Wallops Flight Facility I was forecasting local air temperature and modeling the dependence of lightning frequency on local climate change, working in R on data that did not behave the way the textbooks assume. Working through the residuals turned up something that was not a modeling problem at all: a NOAA sensor was malfunctioning. The forecasting work was published with me as first author, and certain results went into a NASA Technical Memorandum.
That is the thing I would want you to know about how I work. The interesting result usually shows up in the part of the analysis that is not behaving, and most of the value is in taking that seriously instead of smoothing it out.
Tools
Python, R, SQL, and MATLAB day to day. Mathematica and LaTeX for the mathematical work. Haskell, Lisp, and Scala when the problem is functional in shape. Docker, git, and shell for everything around it. Vector, relational, and graph databases. Qiskit and QCL on the quantum side.
Getting started
Email me a description of the problem and, if you can, a sample of the data. I will tell you honestly whether it is a few hours of work or a few weeks, and whether I am the right person for it. If I am not, I will usually be able to say what kind of person is.