Vision — Right Fidelity AI

Why AI, why now

Start with questions the evidence can settle.

AI is useful here when its work can be compared with a lab measurement or a later outcome.

AI can now help propose models, test alternatives, and work through evidence at a scale one researcher could not manage alone. That makes it useful for discovery, provided the result has an external check.

Lab measurements provide that check for variant-effect models. Later clinical outcomes provide it for prognosis. Public data makes some of this work possible now, while restricted clinical studies require separate access, controls, and publication gates.

That boundary matters. Where no measurement or held-out outcome can grade the work, a persuasive answer can still be wrong. I would rather make a narrower claim that can be tested.

Why neurodegeneration

Why begin with neurodegeneration.

This started close to home — the full story is on the motivation page.

The choice is not only personal. Neurodegenerative diseases are among the most feared and least predictable conditions we know because they take memory, speech, movement, and independence. Families need more than a population average or vague reassurance. They need an individual forecast that says what may happen, roughly when, and how uncertain that estimate is.

A forecast can eventually be compared with what actually happens. That makes prognosis both important to families and suitable for the kind of accountable modeling this mission is built around.

The roadmap

A long horizon.

Each stage begins only when the evidence, data access, and partnerships support it. The roadmap separates completed work from active studies and longer-term plans.

Now · Alzheimer’s & ALS

The two I started for

Begin with questions about Alzheimer’s and ALS that can be tested against molecular measurements or later outcomes. Completed evidence and active-study gates are maintained on the Research page.

Next · A repeatable method

“How will this progress?”

Families most often want to know how the disease is likely to progress. The goal is to produce forecasts that state both the expected course and the uncertainty around it, then test the same method in other cohorts.

Later · The wider family

Beyond the first two

If the method holds, it can be tested on Parkinson’s and other neurodegenerative diseases, using molecular measurements, clinical trajectories, and potentially speech.

Work on additional diseases is years away and depends on the earlier stages succeeding. As one researcher, I can build, calibrate, and verify models. I cannot run a clinic or a wet lab.

The plan is to accumulate small, verifiable contributions over years, with each stage depending on the one before it. Pace is secondary to getting the result right.

Standards for publication

Rules for making claims.

These rules apply at every stage and do not change for a better headline.

Publication rules

Report uncertainty. Every prediction includes an interval or another appropriate measure of uncertainty.
Avoid false precision. If the model cannot support a narrow estimate, the published range stays wide or the forecast is withheld.
Tie claims to evidence. Each claim must be checkable against a measurement, a public benchmark, or a held-out outcome.
Publish results and limits together. Both appear in the build log, with the data-handling rules documented separately.

The goal can be ambitious without making the evidence carry more weight than it can bear.

Where this is going

The outcome I am working toward.

The long-term goal is to give families a reliable sense of what may be coming, and roughly when, while there is still time to prepare.

That is narrower than a cure, but it is the thing I kept wishing for. It has to be built and tested one piece at a time.

Read the research → Why this exists →

Pointing AI at the diseases that take who we are.