Frameworks, products and tooling for machine learning, LLM agents, and operational decision systems.

We build the layer between a research result and something a team can run: the project structure, the tracking, the interface, and the tests that keep a number honest six months after it was measured. Everything here is open source, typed, and tested in CI on Linux, macOS and Windows.

Website · RU

Created and maintained by Denis Drobyshev — backend & ML engineer (portfolio).

Answers a question about Russian law and returns the specific articles it rests on, with every citation verified. Retrieval is hybrid — lexical plus dense search with a cross-encoder reranker — and an NLI model checks that each cited norm actually supports the claim, so a plausible-but-wrong reference is caught rather than shipped. The default answer is extractive, the text of the law itself, and cannot hallucinate; an optional LLM layer passes the same check.

Six codes ship in the repository — civil, tax, labour, criminal, housing and administrative offences, the civil one running to 1,712 articles and 4,717 provisions — with the Consumer Protection Act beside them and a cross-reference graph between articles. The texts are transcriptions from Wikisource, marked in the corpus as pending a check against the official publication at pravo.gov.ru; the parser for that source is how an edition gets confirmed rather than how a code gets added.

Judicial practice ships as 148 paragraphs of Supreme Court Plenum rulings, indexed against the articles they construe, so a norm points at the paragraph that interprets it rather than at a ruling as a whole. That set is what Wikisource carries and no more. Mass case law is the part that is still a pipeline rather than a feature: there is no open structured corpus for Russia at the level of Caselaw Access, and kad.arbitr and ГАС «Правосудие» give data up grudgingly.

Retrieval quality is measured by an eval on an eighteen-question golden set. The offline figures are pinned by a test that re-runs it; recall@5 0.92 and MRR 0.94 on the full corpus are measured on a GPU that CI does not have, and are stated here as measurements rather than as guarantees.

docker compose up app        # → http://localhost:8077, no keys required

Python 3.12 · FastAPI · Docker · documentation · releases · Apache-2.0

Django's philosophy, applied to machine learning. You declare datasets, models, agents and evaluations; the framework runs them, versions them, records them and shows them to you. One class body becomes an admin page, a documented API endpoint, a migration and a CLI target at once.

pip install "mlango[sklearn]"
mlango startproject myproject

Agents are a first-class family beside models, sharing one metastore, one admin and one evaluation system. Most tools in this space are for classical ML or for LLMs; most teams are doing both.

Python 3.10+ · documentation · PyPI · MIT

Every model call, tool call and state transition is a plain object you can log, snapshot and replay. No hidden control flow; the whole loop fits in one file you can read in an afternoon.

pip install glia-agents

The crowded agent-framework field has one consistent complaint: too much abstraction, hidden control flow, painful to debug. glia is the opposite bet. It ships the modern techniques — tools, structured outputs, context compaction, durable checkpoints, guardrails, subagents, evals-as-tests — as opt-in primitives you can read, not a monolith you must trust. If you want a graph engine, use LangGraph. If you want a small, transparent loop you fully understand, use this.

Python 3.10+ · zero required dependencies · documentation · PyPI · MIT

Pricing, inventory, energy, queues and supply chains: the decisions a business makes thousands of times a day, where a small policy improvement compounds. Thirty-two algorithms and twenty-four environments — nine of them applied — typed and tested, aimed at problems that have a cost function rather than a leaderboard.

pip install decisionrl

Every applied environment ships with the classical operations-research baseline beside it, so a learned policy is measured against the standard method rather than asserted to be better. Where the classical method is already optimal, the README says so and shows the learned policy matching it.

Python 3.9+ · documentation · PyPI · MIT

Six operational decisions — stock, price, admission to a queue, a battery, two echelons of a supply chain, and pricing and ordering taken together — each with two reference arms beside the agent under test: the classical operations-research method, tuned on seeds held out of the evaluation, and the exact optimum from backward induction.

pip install stadion-rl

The score is normalised between those two and carries a bootstrap interval over paired instances, so "no measurable difference from the classical method" is a verdict the report can return rather than a rounding error it has to hide. The room a task leaves runs from 0.4% on inventory, where the newsvendor formula is already near-optimal, to 26.6% on the battery. That spread is the point: a benchmark whose tasks all leave generous headroom has selected for problems where the textbook answer is bad.

Every other number here is measured against the optimum, so no ordinary run would notice a wrong recurrence. stadion verify computes each dynamic program's value and, separately, simulates the policy that same program emits; the two have to agree within Monte Carlo error, and CI runs it on every push.

Python 3.10+ · PyPI · MIT

A complete roadmap through machine learning, neural networks, reinforcement learning and recommender systems: twenty-seven modules from the arithmetic of a mean to reproducing a recent paper. No sign-up, no first-module-free, no instalments.

The central skill is checking a claim rather than launching a training run. The field moves through papers, most of the improvements they announce do not survive a change of random seed or a comparison against a tuned baseline, and someone who can train a model but cannot check a claim builds on noise. Module 1 is about baselines and confidence intervals, before any machine learning at all.

Written in Russian and in English, both versions complete: twenty-seven modules each, behind the same notebooks. Those run on a CPU in seconds and are executed in CI on Linux and Windows, because a reader whose notebook does not start has no course.

drobyshevdev.github.io/lemma · prose CC BY 4.0 · code MIT

Errors teach. 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.

Tests are named after the guarantee they protect. 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.

Verification beats assertion. A green pipeline on one machine is not evidence. Every claim we publish — an accuracy, a benchmark, a "this works on Windows" — 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.

Comments explain why. The code already says what it does. The comment is for the constraint a reader cannot see.

No hidden control flow. If a framework does something on your behalf, it should be possible to find the line where it happens.

Issues and pull requests are welcome on any project. Every repository has a CONTRIBUTING.md — its own, or the organisation-wide default — with the checks CI will run and the standards review will hold you to. Security reports go through GitHub's private vulnerability reporting rather than a public issue.

Documentation is written in English and Russian, structured so a third language is one file per page rather than a fork.

The organisation is maintained by Denis Drobyshev. The people page lists members who have made their membership public.