• All of it in Julia, with Turing.jl and Mooncake as the autodiff backend
• Our R and Stan tools were not expressive enough
• Nothing in Julia for infectious disease modelling that made sense to use, so all of it from scratch. About 35,000 lines

• The good. Heavy use of submodels makes it easy to follow the workflow
• The bad. Really quite slow. I cannot get Enzyme to work, and Mooncake compile times are very slow
• Some of it took a lot of work to run under autodiff at all, the integrals especially

• Composable tools for infectious disease modelling in Julia
• Models built from interchangeable parts, not one system rebuilt for each outbreak
• A specialist contributes the part they know, without holding the whole model in their head
• Uncertainty carried from the priors through infection dynamics to observation
• Nowcasting, forecasting and scenarios from one model definition

• A library of parts over Turing.jl. Transmission, latent process and observation are separate objects that nest into one model
• Joint inference and the rest of the Turing toolchain come with it
• Registered, and the proof of concept the 16:45 talk demonstrates

• Model the relationships between events with Distributions.jl
• Those relationships form an event tree
• Modify the tree, and the distributions in it, for an intervention. Split the tree, compete distributions against each other
• convolve_series pushes a time series through the tree, and it marginalises across the tree where that is cheaper
• lower turns a delay into a compartmental generator, over Catalyst, SciMLBase, JumpProcesses and AlgebraicPetri

• All of them are used by public health departments, including CDC, UKHSA, RKI, ECDC and WHO
• epinowcast was the attempt at community developed tools
• In many cases the package is too complex for anyone else to use
• We see fragmentation, and a lot of small mistakes in outbreak modelling that add up to real biases

• I was travelling for summer schools, teaching and workshops, so decided to have a go with agents
• 966 merged pull requests in the EpiAware org from the bot account, and 119 closed unmerged
• A Claude Max subscription
• Well, actually, we still need funding

• Every package’s workflows are thin callers into 22 reusable workflows in one .github repository, as SciML does it
• Each caller pins the reusable workflow to a commit SHA
• The more interesting ones: the per-backend autodiff matrix, downstream testing, and the weekly template sync
• release-nudge opens a tracking issue when a package has unreleased changes, and closes it again when it does not

# MANAGED by EpiAwarePackageTools.scaffold
# — do not edit by hand.
name: Test

jobs:
  test:
    uses: EpiAware/.github/.github/workflows/tests.yml@d024dc6

try-this-pr comments install instructions on every pull request

import Pkg
Pkg.activate(temp = true)
Pkg.add(url = "...", rev = "my-branch")
using CensoredDistributions

• Rigid template implementing and enforcing community standards
• scaffold writes the tooling into a package once. update re-applies it and reports drift
• Aqua, JET, formatting, docstrings and benchmarks
• Sixteen written standards, and checks for them
• It injects AGENTS.md and CLAUDE.md so a model can find the docs

• CRAN has more checks than comparable package servers
• As part of this there is a set of automated checks that looks at documentation, examples, dependencies and portability
• In Julia you can opt in to Aqua, which is great
• But we want to go further, so we added JET, docstring format, doctests, README sections, import centralisation, formatting and benchmarks

• We do not think there is a community standard for what supporting automatic differentiation means
• So we aimed to build our own, using DifferentiationInterfaceTest
• In SpecialFunctions.jl the shape-parameter partials of gamma_inc and beta_inc are not implemented, so support is janky across backends
• EpiAwareADTools.jl stores our org-wide automatic differentiation fixes, and aims to delete each one when upstream ships it

• Julia’s default is a docstring if you feel like it
• DocStringExtensions gets us close to R’s standard. I would like more
• The README’s Julia blocks run at docs build, so a broken example fails the build
• A tutorial for each key feature, developed alongside the feature and checked in CI
• SciML puts something interesting on the front page and keeps the tutorials short. We copy that

• Nobody from outside the team has committed since February, so the other party on an issue is usually one of my robots. Answering a robot is dispiriting
• It is not clear which repositories want robot-filed issues at all
• Do 🤖 shut down collaboration?
• EpiAwareAgents is one possible answer, a community bot any maintainer could point at their own repository
• 🌍 We want contributors who are not Julia developers, and not in well resourced settings

• epiaware.org — the packages, the docs, and how to get involved
• github.com/EpiAware — including EpiAwarePackageTools.jl and EpiAwareADTools.jl
• BVDOutbreakSize — the live DRC outbreak size model
• A workflow for infectious disease modelling — the workflow we are trying to follow
• The prompts that wrote these decks