Apppublic

SameerInfinity/AI_DD

sourceHugging Faceupdated 6mo agoView on Hugging Face

0likes

App README

ABYSS — AI-Powered M&A Due Diligence Virtual Data Room

A multi-party Virtual Data Room (VDR) platform for mergers & acquisitions due diligence. Upload documents, get perspective-specific AI analysis with structured findings, risk scores, and source citations. Supports Buy Side, Sell Side, and Admin workflows with role-based access, request lists, audit trails, and Human-in-the-Loop review.

Live: Frontend (Vercel) | Backend (HF Spaces)

Architecture

┌─────────────────────────────────────────────────────────────────┐
│                FRONTEND (React 18 + TypeScript + Vite 8)        │
│  Auth → Perspective → Projects → Data Room / Q&A / Audit       │
│  UI: shadcn/ui + Tailwind  |  State: Zustand + React Query     │
│  Theme: Ocean dark + Sunny light  |  Code-split (React.lazy)   │
└─────────────────────────┬───────────────────────────────────────┘
                          │ REST API (JWT Auth)
┌─────────────────────────▼───────────────────────────────────────┐
│                   BACKEND (FastAPI + async)                      │
│                                                                  │
│  Auth ─ Projects ─ Teams ─ Documents ─ Requests ─ Queries       │
│  Folders ─ DD Tests ─ Routines ─ Planning ─ Finance ─ Export    │
│  Knowledge ─ Self-Improvement ─ VDR Security                    │
│                                                                  │
│  ┌────────────────────────────────────────────────────────────┐  │
│  │         OCR → Chunker → Embeddings → Vector Store          │  │
│  │    (PyMuPDF / Tesseract → Semantic Split → Pinecone EU)    │  │
│  └────────────────────────────────────────────────────────────┘  │
│                                                                  │
│  ┌────────────────────────────────────────────────────────────┐  │
│  │              AI PIPELINE (8-Step State Machine)             │  │
│  │  Rewrite → Classify → Retrieve → Completeness              │  │
│  │  → Think (Groq) → Audit → Structure → Done                 │  │
│  │  + Review feedback injected into system prompt              │  │
│  └────────────────────────────────────────────────────────────┘  │
└──────────────────────────────────────────────────────────────────┘
         │                │                 │
   ┌─────▼──────┐  ┌──────▼───────┐  ┌─────▼──────────┐
   │ Supabase   │  │   Pinecone   │  │ Supabase       │
   │ PostgreSQL │  │  (vectors)   │  │ Storage (files) │
   │  (EU)      │  │  (EU)        │  │  (EU)          │
   └────────────┘  └──────────────┘  └────────────────┘

Tech Stack

Layer	Technology
Frontend	React 18.3 + TypeScript + Vite 8 (SWC), code-split via React.lazy
UI	shadcn/ui + Tailwind CSS + Radix UI + Framer Motion
State	Zustand 5 + TanStack React Query 5 (5-min staleTime cache)
Forms	React Hook Form 7 + Zod validation
Theme	Dark (ocean abyss) + Light (sunny ocean) with smooth 400ms transition
Export	jsPDF + html2canvas (PDF), python-docx (DOCX), openpyxl (XLSX)
Backend	FastAPI 0.115 + Uvicorn (fully async)
Database	Supabase PostgreSQL EU (SQLAlchemy AsyncSession + asyncpg)
File Storage	Supabase Storage EU (with local fallback)
LLM	Groq — Llama 3.3 70B Versatile
Embeddings	sentence-transformers (all-MiniLM-L6-v2), 384-dim, local
Vector DB	Pinecone serverless EU (local in-memory fallback)
OCR	PyMuPDF + Tesseract (digital + scanned PDFs, images)
Auth	JWT (python-jose) + bcrypt + Google OAuth + email verification
Email	Brevo (verification, password reset, project invitations)

All 11 Modules

1. Auth Module

—Register, login, Google OAuth, email verification (Brevo), password reset
—8 analyst perspectives (set once, changeable from sidebar)
—5 project roles: Admin, Buy Side (Legal/Tax/Commercial), Sell Side
—Email-based project invitations (admin sends, user accepts/declines)
—Session management with global 401 redirect

2. Project Module

—Create, rename, delete projects with extended fields (legal form, deal type, country, employees, revenue, deal value)
—Email invitation system (replaced invite codes)
—Team management with role badges and activity feed
—VDR security settings (watermarks, download restrictions per role)

3. DMS Module (Data Room)

—Upload PDFs, DOCX, XLSX, CSV, images, plain text
—Nested folder structure with directory upload
—Document versioning, AI auto-tagging, full-text search
—Professional list-view file manager
—Language detection (AI responds in document language)
—Document watermarks (username + date overlay, configurable)

4. OCR Module

—PyMuPDF for digital PDFs, Tesseract for scanned documents
—python-docx for Word, pandas + openpyxl for Excel/CSV
—Semantic chunking (1200 chars, 150 overlap) → Pinecone indexing
—Reprocess-all endpoint for failed documents

5. Agent Module (AI Pipeline)

—8-step pipeline: Rewrite → Classify → Retrieve → Completeness → Think → Audit → Structure → Complete
—8 perspective-specific system prompts
—Citation enforcement, confidence scoring, red flag detection
—Human-in-the-Loop review (Approve/Edit/Reject) on all AI outputs
—Review feedback injected into system prompt (AI learns from corrections)

6. Planning Module

—Interactive AI-guided audit planning dialog
—AI analyzes project profile, asks 4-6 targeted risk questions
—Auto-generates DD request list by workstream
—Sessions persisted — visible to all team members
—Past sessions viewable in read-only mode

7. Finance Module

—Upload accounting data (Excel/CSV), auto-detect SKR 03/SKR 04
—Range-based account mapping (100+ German standard account ranges)
—Journal entry format support (debit/credit per row)
—Balance sheet items excluded from P&L automatically
—Variance analysis: Month-over-Month, Year-over-Year, anomaly detection (>30%)
—Key metrics: EBITDA margin, gross margin, net margin, personnel ratio, equity ratio, ROA

8. Report Module

—PDF export: Q&A analysis with structured findings, TOC, confidence bars
—DOCX export: Full DD report with request list, test results, Q&A summary
—XLSX export: Workstream sheets with status, scores, findings
—Editable executive summary (write or AI-generate before export)
—Download restrictions enforceable per role

9. Knowledge Module

—20 built-in IDW/M&A standards (Legal, Tax, Commercial, General)
—Auto-seeded on first access, filterable by category
—AI-generated cross-project insights ("Generate Insights" button)
—User-level, accessible from Projects page

10. Audit Module

—Completeness audit: AI cross-checks documents against request list
—Per-field DD tests with pass/fail/warning scoring
—Custom test routines with user-defined prompts
—Results persisted — load on revisit, shows "Last run by [name]"
—Activity log: human-readable action labels, user/IP tracking
—Cookie consent + GDPR privacy policy

11. Self-Improvement Module

—Usage analytics: projects, queries, uploads, tests, reviews
—30-day activity heatmap
—AI review patterns: approve/edit/reject rates with contextual messages
—Actionable recommendations: missing docs, open requests, unrun audits
—User-level, accessible from Projects page

Performance

Optimization	Effect
Route-based code splitting	Main bundle 970KB → 361KB (63% smaller). Each page loads on-demand.
Hover preloading	Page chunks preload on sidebar hover — zero blink on first visit.
React Query cache	5-min staleTime, no refetch on tab switch. Data survives page navigation.
Splash skip	Repeat visits within session skip 4s splash animation.
Smart polling	Exponential backoff (2s → 30s), stops when idle. Zero requests when nothing processing.
Theme transition	Temporary 400ms CSS transition class — only color/bg/border, no layout cost.

Pages (21 total)

Page	Route	Description
Sign In	`/signin`	Email/password + Google OAuth
Sign Up	`/signup`	Registration with password strength checklist
Verify Pending	`/verify-pending`	Email verification waiting page with resend
Verify Email	`/verify-email`	Token verification from email link
Forgot Password	`/forgot-password`	Request password reset email
Reset Password	`/reset-password`	Set new password with strength checklist
Perspective	`/role`	8 analyst perspectives (shown once after signup)
Projects	`/projects`	Create projects, accept/decline invitations
Knowledge	`/knowledge`	DD standards + AI cross-project insights
Insights	`/insights`	Usage analytics + AI review patterns
Q&A	`/dashboard`	AI query interface with branching conversations
Planning	`/planning`	Interactive audit planning with session history
Requests	`/requests`	Request list with workstream tabs + per-item upload
Data Room	`/dataroom`	File manager with folders, search, versions, watermarks
Finance	`/finance`	Accounting import, variance analysis, key metrics
Audit	`/audit`	Completeness audit, field tests, custom routines
Upload	`/upload`	Drag & drop upload with progress tracking
Activity	`/activity`	Full audit trail with human-readable actions
Team	`/team`	Members, invitations, VDR security settings, activity feed
Privacy	`/privacy`	GDPR privacy policy
404	`*`	Not found page

Database Schema (19 Tables)

Table	Purpose
users	Accounts with perspective (8 types), email verification
projects	Deal containers with extended fields, VDR security settings
project_members	Per-project role assignments (admin/buy/sell)
project_invites	Email invitations with accept/decline status
documents	Files with versioning, tags, language, OCR status
chunks	Indexed text segments with Pinecone IDs + page numbers
folders	Nested folder structure for data room
conversations	Chat thread branches per project/user
queries	Questions + pipeline state + Human-in-the-Loop review
pipeline_steps	Step-by-step tracking with I/O + duration
request_items	DD request list items with workstream + status
dd_test_results	Due diligence test results with findings
test_routines	User-defined custom test prompts
audit_logs	Full activity trail: who, what, when, IP
audit_results	Saved completeness audit results with history
planning_sessions	Persisted planning conversations
financial_datasets	Uploaded accounting files metadata
financial_entries	Parsed accounting line items
knowledge_entries	DD standards + AI-generated insights

Enums: UserRole (5), Perspective (8), Workstream (4), InviteStatus (4), DocStatus (4), QueryStatus (9), RequestStatus (5), Severity (5)

Environment Variables

Variable	Required	Description
`DATABASE_URL`	Yes	PostgreSQL connection (async: `postgresql+asyncpg://...`)
`JWT_SECRET`	Yes	Secret key for JWT tokens
`GROQ_API_KEY`	Yes	Groq API key for Llama 3.3 70B
`SUPABASE_URL`	Yes	Supabase project URL
`SUPABASE_SERVICE_KEY`	Yes	Supabase service role key
`BREVO_API_KEY`	Yes	Brevo API key for transactional emails
`BREVO_FROM_EMAIL`	Yes	Verified sender email address
`FRONTEND_URL`	Yes	Frontend URL for email links
`PINECONE_API_KEY`	No	Pinecone API key (falls back to local)
`PINECONE_REGION`	No	Pinecone region (default: `eu-west-1`)
`GOOGLE_CLIENT_ID`	No	Google OAuth client ID
`EMBEDDING_MODEL`	No	Embedding model (default: `all-MiniLM-L6-v2`)

Quick Start

bash

# 1. Clone
git clone https://github.com/SameerInfinity/AI-Due-Diligence.git
cd AI-Due-Diligence

# 2. Backend
python -m venv .venv
source .venv/bin/activate        # Windows: .venv\Scripts\activate
pip install -r requirements.txt

# 3. Frontend
cd frontend && npm install && cd ..

# 4. Configure
cp .env.example .env
# Edit .env with your keys

# 5. Run backend
uvicorn backend.main:app --reload --port 8000

# 6. Run frontend (new terminal)
cd frontend && npm run dev

Deployment

Service	Platform	Region
Backend	HuggingFace Spaces (Docker)	EU
Frontend	Vercel (Edge CDN)	Global
Database	Supabase PostgreSQL	EU West (Paris)
File Storage	Supabase Storage	EU
Vector DB	Pinecone	EU West (Ireland)
Email	Brevo	Global
Redis	Upstash	EU