ADM Capital � ESG Diligence & ESAP Tracker

An AI workflow layer that sits inside ADM Capital's Environmental & Social Management System (ESMS). Borrower documents go in; an IFC Performance Standard-structured ESG risk memo and an evidence-linked, draft ESAP come out, then flow into a KPI monitoring view.

Built around a single fictional borrower � PT Sumber Laut Lestari (SLL), an Indonesian vannamei shrimp producer � using the real five-document demo pack in the parent folder.

The demo loop (?5 minutes)

1. Intake � ingest SLL's mixed-format pack (Word / Excel / PDF); auto-classify and extract deal entities.
2. Risk memo � generate a memo organised by IFC PS1�PS8 + Climate, ~13 findings with severity, each citing its source passage across multiple documents, plus a missing-information list. Every field is editable.
3. ESAP � convert findings into an IFC-PS-tagged action plan (owner, due date, required evidence, status, condition precedent / subsequent). Export to Word and PDF; one-click borrower-friendly summary.
4. Monitor � seeded KPI dashboard with traffic-light ESAP/KPI status and SFDR PAI references (fa�ade).
5. Portfolio � aggregate risk by IFC PS / sector / country and the recurring pattern (fa�ade).

Run it

npm install
npm run dev      # web on :5173, API on :8787

Open http://localhost:5173.

Live LLM (optional)

The core "messy docs ? IFC-PS risk memo" engine runs live on Anthropic when a key is present, and otherwise falls back to bulletproof pre-baked SLL results � so the demo never stalls.

cp .env.example .env
# add ANTHROPIC_API_KEY=...   (optionally ANTHROPIC_MODEL=)
npm run dev

The header badge shows Live LLM vs Demo engine.

Architecture

• Demo build (this repo): Vite + React + TypeScript + Tailwind front end; a thin Express proxy to the LLM; in-memory state; client-side Word/PDF export. No DB, queue, OCR or auth.
• Production vision: Next.js + FastAPI; PostgreSQL (records / tasks / KPIs / audit logs); Redis + Celery/Temporal for document processing; OCR (Textract / Azure Doc Intelligence); RAG over ADM policies, IFC PS, SFDR PAI and past ESAPs; S3; RBAC; Docker.
• Data residency: this prototype runs on a frontier LLM; in production for a Hong Kong-based manager it would run through Amazon Bedrock or Google Vertex AI with regional data residency, keeping borrower data within an approved boundary.

Design principles

Mirror the ESMS � IFC PS native � human-in-the-loop (every output is a draft the analyst owns) � evidence-linked / audit-ready � speaks ADM's language � exports into existing artefacts. PT Sumber Laut Lestari is fictional, created for demonstration.