Esgard

CSRD/ESRS compliance copilot — document-first, evidence-traced, audit-ready. Runs 100% on the free Hugging Face tier.

Open datapoint library How it works Benchmark

What it does

Esgard turns the annual reports, policies and data exports a company already holds into audit-ready EU CSRD/ESRS disclosures. Every extracted datapoint binds to a verbatim quote, chunk and page number — the evidence trail assurance requires.

Ingest — PDF/DOCX/TXT into heading-aware chunks with page/line provenance (PyMuPDF, no GPU).
Map — content to a curated, versioned ESRS datapoint library (derived from Reg. (EU) 2023/2772; revised-ESRS FY2027 upgrade on the roadmap).
Extract — deterministic, unit-aware extraction; optional LLM enhancement (Qwen2.5-7B-Instruct) with quote re-verification.
Analyse — four-state coverage per datapoint: evidenced / partial / absent / needs_input.
Draft — ESRS-structured report skeleton with inline citations; FastAPI + Gradio front ends.

Benchmark (v0.1 evaluation set)

Metric	Value
Precision	1.0
Recall	1.0
F1	1.0
Evidence rate	1.0
Cases	12 (E1, S1, G1 numeric / percentage / boolean)

Product status

Working pipeline (ingest → retrieve → extract → gap analysis → draft) with 41 passing tests, ruff + mypy clean.
~70 curated datapoints published openly as hardik90/esgard-esrs-datapoints.
Business plan, pitch deck, architecture and research in the repository docs.

Run it yourself

git clone <repo> && pip install -e ".[dev]"
python -m pytest                     # 41 tests
python scripts/run_benchmark.py      # eval set
uvicorn esgard.api.app:app --reload  # REST API at /docs

Esgard is a compliance copilot, not legal or assurance advice. MIT licensed; datapoint library derived from EU public legal texts.