🛡️ CriteriaGuard: Governance-Grade AI for Procurement

Explainable AI Platform for Indian Government Tender Eligibility Evaluation

CriteriaGuard is a high-integrity, end-to-end platform designed to automate the manual, error-prone process of cross-checking bidder submissions against tender eligibility criteria. Built specifically for the complexities of Indian Government Procurement, it ensures every decision is deterministic, traceable, and fully auditable.

---

🏛️ The Problem: The "Governance Gap"

Every year, procurement committees spend days manually reviewing 100+ page tender documents and diverse bidder submissions (scanned certificates, typed PDFs, photographs).

• The Risk: A single missed condition leads to a wrongful award or a court challenge.
• The Transparency Gap: Manual decisions are often untraceable, inviting RTI inquiries.
• The AI Trap: Generic AI "black boxes" can hallucinate evidence, which is unacceptable for high-stakes government work.

---

🚀 The Solution: CriteriaGuard

CriteriaGuard doesn't replace the procurement officer; it augments their expertise with a consistent, evidence-backed first-pass evaluation.

🛡️ Core Pillars

1. Explainable AI (XAI): Every "Eligible" or "Not Eligible" verdict is backed by a direct citation (Document Name, Page Number, and Excerpt).
2. Deterministic Logic: AI performs the extraction, but pure Python code (VerdictCore) performs the evaluation. No hallucinations in the final decision.
3. Tamper-Evident Logs: SHA-256 chained audit logs ensure that no evaluation result can be quietly altered.
4. Human-in-the-Loop: High-ambiguity clauses and borderline numeric values are automatically routed to a human officer for sign-off.

---

🏗️ System Architecture (The 3-Stage Pipeline)

graph TD
    A[Tender Document] -->|CriteriaLens| B(Stage 1: Tender Intelligence)
    B -->|Structured Schema| C{Officer Approval}
    C -->|Verified Criteria| D(Stage 2: Bidder Understanding)
    E[Bidder Submissions] -->|DocProbe| D
    D -->|Extracted Values + Citations| F(Stage 3: Verdict Engine)
    F -->|Deterministic Rules| G[Audit-Ready Dashboard]
    G --> H[Signed PDF Report]
    
    subgraph "The Intelligence Layer"
        B1[Llama 3.3 70B via Groq]
        B2[Ambiguity Resolver]
        B1 --- B2
    end
    
    subgraph "The Extraction Layer"
        D1[Multi-Format OCR]
        D2[Layout Preservation]
        D1 --- D2
    end

Loading

---

🛠️ Features

1. Stage 1 — CriteriaLens (Tender Intelligence)

• Extracts criteria into a formal schema: technical, financial, and compliance.
• Linguistic Marker Analysis: Differentiates between mandatory ("shall", "must") and optional ("should", "preferred") clauses.
• Officer Checkpoint: Provides a clean checkpoint for the procurement officer to approve the extracted requirements before evaluation begins.

2. Stage 2 — DocProbe (Bidder Understanding)

• Multi-format support: Direct PDF extraction, OCR for scanned documents, and Word (.docx) support.
• Contextual Anchoring: Locates the exact paragraph and value, recording the source reference.
• Authenticity Scoring: Evaluates the quality of the source document to flag low-confidence extractions.

3. Stage 3 — VerdictCore (Explainable Verdicts)

• Zero-Hallucination Engine: Final verdicts are computed via pure deterministic logic.
• Borderline Detection: Automatically flags numeric values within 10% of a threshold (e.g., if turnover is ₹4.9Cr against a ₹5Cr requirement) for human review.
• Needs Review Queue: Routes any low-confidence or ambiguous case to a human expert with a plain-English explanation of why the system is unsure.

4. Integrity Suite

• Tamper-Evident Audit Trail: Append-only log with SHA-256 chaining, suitable for formal record-keeping.
• Governance Reports: Generates signed, audit-ready PDF reports with full citation tables for every bidder.

---

💻 Technology Stack

• Backend: FastAPI (Python 3.11), Pydantic v2 (Schema Validation).
• CriteriaGuard Frontend: React 18 (Vite), Glassmorphism UI, High-Performance WebGL (Aurora) animations.
• LLM Layer: Llama 3.3 70B (Groq) for high-speed, accurate extraction.
• OCR Engine: Tesseract & Cloud Vision Ensemble.
• Database: PostgreSQL (Supabase) with SHA-256 Chaining.
• Deployment: Containerized (Docker ready) for NIC/MeitY-approved infrastructure.

---

🚦 Getting Started

Prerequisites

• Python 3.11+
• Node.js 18+
• Groq API Key (for Llama 3)
• Supabase Credentials

Quick Start

1. Clone the Project

   git clone https://github.com/Saksham-official/CriteriaGuard.git
   cd CriteriaGuard

2. Backend Setup

   cd backend
   python -m venv venv
   source venv/bin/activate # Windows: .\venv\Scripts\activate
   pip install -r requirements.txt
   # Setup .env with GROQ_API_KEY and SUPABASE_URL/KEY
   python main.py

3. CriteriaGuard Frontend Setup

   cd frontend
   npm install
   npm run dev

---

📜 Governance Commitment

CriteriaGuard is designed to be domain-agnostic. Whether it is defense (CRPF), infrastructure, or health, the system adapts to any tender structure. It sits behind the existing process, making it faster, more consistent, and 100% traceable.

---

“Built for the realities of Indian Government Procurement—where accountability meets intelligence.” 🛡️