Frameworks and tooling for machine learning, LLM agents and operational decisions.

Answers a question about Russian law and returns the specific provisions it rests on, with every citation verified. Retrieval is hybrid — BM25 plus dense embeddings 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.

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

Every model call, tool call and state transition is a plain object you can log, snapshot and replay. The whole loop fits in one file you can read in an afternoon. The modern techniques — tools, structured outputs, context compaction, durable checkpoints, guardrails, subagents, evals-as-tests — ship as opt-in primitives you can read, not a monolith you have to trust. If you want a graph engine, use LangGraph; if you want a small, transparent loop you fully understand, use this.

Pricing, inventory, energy, queues and supply chains: the decisions a business makes thousands of times a day, where a small policy improvement compounds. 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 good — and where the classical method is already optimal, the README says so.