Certo ESG Compliance Agent
Local AI assistant for CSRD/ESRS gap analysis and compliance evidence
A compliance assistant that checks company sustainability documents against the 10 ESRS standards and answers questions with a locally run language model, so documents are not sent to an external AI service. Later versions added an evidence graph, audit logging and a desktop app for triaging regulatory alerts. Work has been paused at the prototype stage since February 2026.
Key points
- Report rates disclosure for each of the 10 ESRS standards as complete, partial or missing, with a 0–100 total
- Answers questions with a local 4-bit Granite 4.0 Tiny model and FAISS retrieval; no external AI APIs called
- Evidence graph links each requirement to controls, evidence and IT systems, and reports coverage gaps
- Triggers block edits and deletes in the main audit table; Radar events are written to a hash-chained log
What it is
Certo ESG Compliance Agent checks a company's sustainability reporting against the European Sustainability Reporting Standards (ESRS) required under CSRD. It reads sustainability reports and related documents, flags standards with missing or partial disclosures, and produces an HTML report that lists the supporting documents for each standard. Questions are answered by a language model running on local hardware, so confidential documents can be analysed without sending them to an external AI service. The intended users are large and mid-cap companies in Poland and the EU that must report under CSRD.
How it works
Documents are split into chunks, embedded with BGE and indexed in FAISS for semantic search. Retrieved passages go to a Granite 4.0 Tiny model, quantised to 4 bits and run locally through llama.cpp, which answers questions in five context modes: ESRS, general ESG, regulation, gap analysis and summarisation. The HTML report applies fixed rules: each standard is rated complete, partial or missing depending on how many matching documents are found, and the ratings are averaged into a 0–100 score. The documentation lists 26 data sources, including EUR-Lex, ESMA, company reports, market data and climate data.
- E1–E5: climate change, pollution, water and marine resources, biodiversity, circular economy
- S1–S4: own workforce, value-chain workers, affected communities, consumers and end-users
- G1: business conduct
Evidence graph and Regulatory Radar
Later versions added two layers. A relational evidence graph links each regulatory requirement to the controls that address it, the evidence behind those controls and the IT systems in scope, and reports coverage gaps. Regulatory Radar sorts incoming alerts into dismiss, review or escalate, with a short note from the local model on why each one matters; accepted alerts become tasks with an owner and a deadline, exportable as a compliance dossier. The same workflow runs in an offline Electron desktop app. Alerts still come from sample data.
Status
Development has been paused since February 2026; the last commit on the main branch landed on 24 February 2026, and the repository is private. The CSRD/ESRS pack is the most developed; NIS2 is in beta, and DORA and the AI Act remain experimental. The code is a pre-pilot prototype: security hardening and a customer pilot would come before any deployment.
Record
- Status
- Paused
- Stage
- Paused, pre-pilot prototype
- Sector
- RegTech / ESG compliance
- As of
Key figures
- ESRS standards assessed
- 10
- Question-answering modes
- 5
Topics
RegTech, CSRD / ESRS, RAG, Local LLM, Python
Questions about this project?
Contact the founders