Open source · Python · MIT and Apache-2.0
The layer between a research result and something a team can run: the project structure, the tracking, the interfaces, and the tests that keep a number honest six months after it was measured.
Five libraries, a free course and open reading notes. Since 2026.
Projects
Each one installs today, carries documentation and is released under a permissive licence. The numbers beside each project come from a script in the repository that produced them.
A legal assistant whose citations are checked, not asserted. Hybrid retrieval with a cross-encoder reranker, and an NLI model that verifies each cited provision supports the claim. The default answer is the text of the law and cannot hallucinate.
Django's philosophy applied to machine learning. You declare datasets, models, agents and evaluations; one class body becomes an admin page, a documented endpoint, a migration and a CLI target at once.
An agent loop you can read in an afternoon. Every model call, tool call and state transition is a plain object you can log, snapshot and replay. The core imports nothing outside the standard library.
Reinforcement learning for decisions that have a cost function: pricing, inventory, energy, queues, supply chains. Every applied environment ships the classical operations-research baseline beside it.
A proving ground where an agent is scored against the exact optimum. Six operational decisions, each with the tuned classical method and the computed optimum beside it, and a bootstrap interval on the difference.
Stack
A short list on purpose. Every dependency is something a reader has to install, learn and keep updated, so each one here earns its place in more than one project.
Typed throughout, checked with mypy and linted with ruff in CI. The version floors are lower bounds on what a user may already have, not the version CI runs — raising them would lock people out without testing anything new.
Every algorithm in decisionrl. CPU is the default target: the applied environments are small enough that a reader without a GPU can still reproduce the numbers in the documentation, which is the point of publishing them.
One API core serves every client. Dense retrieval is a matrix-vector product over vectors held in memory and cached to disk, with a pure standard-library fallback so the offline image and CI run without NumPy at all. A Postgres index with pgvector is designed and not yet wired — the schema is in the repository, the retriever is not.
The image installs the API extra, sets offline mode and carries the corpus, so
docker compose up app runs with no key, no GPU and no outbound request.
A reader who cannot start it has no way to check anything.
Linux, macOS and Windows on every push, with one aggregate check to require in branch protection. Releases publish through Trusted Publishing, and the desktop binaries carry a signed provenance attestation.
Built with --strict, which fails on a broken internal link or a page
missing from the navigation. No webfonts and no analytics on any of the sites,
including this one.
Evidence
There are no client logos here and no case studies. What an open-source organisation can show instead is its measurements, and the commands that reproduce them.
| Task | Headroom | 95% interval |
|---|---|---|
inventory | +0.4% | [+0.561, +1.031] |
joint-pricing | +4.3% | [+3.794, +5.656] |
pricing | +6.2% | [+1.265, +1.761] |
queueing | +16.9% | [+3.418, +3.988] |
supply-chain | +17.3% | [+5.505, +7.493] |
energy | +26.6% | [+4.221, +4.690] |
The spread is the point. Where the textbook answer is already near-optimal there is almost nothing to win; where it cannot plan ahead, a quarter of the value is still on the table. A benchmark whose tasks all leave generous headroom has selected for problems the standard method happens to be bad at.
Notes and teaching
A paper says what it found. A note says what it showed, which is a smaller thing, and marks the distance between them.
Midbrain dopamine neurons fire in a pattern that looks like the error term of temporal-difference learning. Close enough that an algorithm from 1988 predicted what an electrode recorded in 1997 — which is why the result gets overstated.
Machine learning, reinforcement learning, neuroscience and the methods used to check a claim. Every note declares the sources it rests on, and CI rejects one that does not.
From the arithmetic of a mean to reproducing a recent paper. No sign-up and no first module free. Module one is about baselines and confidence intervals, before any machine learning at all, because the central skill is checking a claim rather than launching a training run.
Method
Review criteria, not aspirations. A pull request that breaks one of them gets comments.
A message is read at the worst possible moment by someone who does not have the source open. It should say what went wrong and what to do next, and list the alternatives when there are any.
test_assignment_is_stable_when_rows_are_added, not test_split.
A test whose name does not survive being read aloud is not documenting anything.
A green pipeline on one machine is not evidence. Every published claim is measured on the run it describes, and the numbers in a README are pinned by tests, because prose rots quietly and a failing test does not.
The code already says what it does. The comment is for the constraint a reader cannot see.
If a framework does something on your behalf, it should be possible to find the line where it happens.
Working together
DrobyshevDev
CONTRIBUTING.md — its own or the organisation-wide default — with the checks CI will run, so there is no unwritten rule to discover afterwards.