I spent a while building ESG reporting tooling, and the same thing kept happening: people would paste their internal policies, ledgers and certificates into a chatbot, get back something fluent, and have no way to tell which parts were grounded in their own documents and which the model had invented. Two things in here are my attempt at fixing that. First, the agents that read your files never write report prose. One agent reads each document into a summary; another decides which materials a given chapter may draw on. The actual drafting happens separately, block by block, and each block only sees the facts assigned to it. So "don't make things up" is a structural property, not a line in a prompt. Second, nothing ships until a deterministic gate says so — plain code, no model, checking completeness and consistency against the disclosure standard. Same report, same verdict, every time. It runs entirely on your own machine: Postgres and auth in local Docker, your files on your own disk, and you bring your own model key (Azure OpenAI by default, but any OpenAI-compatible endpoint works — including a local Ollama). Three rule sets ship today: Shanghai Stock Exchange, and HKEX in both Traditional Chinese and English. There's a fully synthetic corpus in the repo, so you can run the whole pipeline end to end — all the way to a rendered Word file — without a model call or any data of your own. MIT, no paid tier, no feature gates. It drafts; it does not assure — what comes out is a draft for professional review. Happy to go into the architecture if anyone's curious, especially the provenance panel — every generated paragraph can tell you which file and which of your answers it drew on.

Sustainability WorkBench

Self-hosted AI workbench for drafting ESG reports

Self-hosted AI workbench for drafting ESG reports