💎 PRISM AI

Offline-First Multimodal Content Transformation Platform

Prism AI allows users to input prompts, attach source documents (PDF, DOCX, TXT, images, audio), configure precise parameters, and instantly generate, preview, and export content across multiple formats (DOCX, PPTX, PDF, Images, Video Scripts).

---

📑 Table of Contents

• 🌟 Key Platform Features
• 🏗️ Architecture & Data Flow
• 🚀 Quick Start & Installation
  • 1. Infrastructure (Docker)
  • 2. Backend (FastAPI)
  • 3. Frontend (React/Vite)
• 💻 Usage & Workflows
• 📂 Project Directory Structure
• 🔒 Security & Privacy Statement

---

🌟 Key Platform Features

• 100% Offline Privacy & Speed: Operates entirely locally with zero external data transmission for maximum security. Seamlessly scalable to cloud deployments when requested.
• Task-Specialized Local Fine-Tuned LLMs: Routes lighter tasks to smaller, optimized models to minimize power consumption, optimize compute efficiency, and accelerate generation times.
• Few-Shot Template Engine: Accepts custom user templates and formatting guides to produce output matching exact organizational standards.
• End-to-End Input & Output Guardrails:
  • Input Guardrails: Presidio PII redaction, file format validation, payload size enforcement.
  • Output Guardrails: Sentence-transformers hallucination & grounding verification, Detoxify toxicity analysis, RapidFuzz plagiarism detection, and schema/format validation.

---

📸 Interface Preview

Step 1: Prompt Input

Step 2: Attach Source

Step 3: Select Parameters

Step 4: Output Screen

Step 5: Template Given & Generated Hindi Output

---

🏗️ Architecture & Data Flow

3-Database Core Infrastructure

1. PostgreSQL: Stores authenticated user data, user sessions, job parameters, and prompt execution history.
2. Qdrant Vector DB: Vector store for document chunks, semantic embeddings, and high-precision RAG contextual retrieval.
3. MinIO Object Storage: Stores uploaded source files (PDF, DOCX, media) and generated output artefacts.
4. Redis: Caching layer, Celery background task queue broker, and real-time state engine.

End-to-End Pipeline

graph LR
    A[Vite/React UI] -->|REST API| B(FastAPI Backend)
    B -->|Task Queue| C{Redis + Celery}
    B <-->|Auth & Metadata| D[(PostgreSQL)]
    B <-->|RAG Embeddings| E[(Qdrant DB)]
    B <-->|File Storage| F[(MinIO Vault)]

Loading

Step-by-Step Flow:

1. Input Phase: User inputs a prompt, uploads source files, and selects target formats.
2. Guardrail Check: Input passes through Presidio PII anonymizer and MIME type validation.
3. Chunking & RAG Retrieval: Source text is extracted, chunked, embedded via sentence-transformers, and indexed in Qdrant.
4. Compute Routing & LLM Generation: Prompt classifier determines task complexity and routes to quantized local fine-tuned LLMs.
5. Output Guardrail Validation: Generated text is scored for grounding, toxicity, plagiarism, and format compliance (with SHA-256 integrity hash).
6. Export: Formatted files (DOCX, PPTX, PDF, Video Script) are saved to MinIO and served via presigned download links.

---

🚀 Quick Start & Installation

To run the full platform locally, follow these 3 setup phases.

1. Infrastructure (Docker)

Navigate to the source code folder and spin up the databases:

cd "1. Source_code"

# Start PostgreSQL, Qdrant, MinIO, and Redis in detached mode
docker-compose up -d

# Verify services are running
docker-compose ps

2. Backend (FastAPI)

Initialize the Python backend server:

cd "1. Source_code/backend"

# Create and activate virtual environment
python -m venv venv
.\venv\Scripts\Activate.ps1    # Windows
# source venv/bin/activate     # Linux/Mac

# Install dependencies
pip install -r requirements.txt

# Start FastAPI Backend Server
python main.py

Note: The backend will run on http://localhost:8000 with interactive Swagger docs at http://localhost:8000/docs.

(Optional) Start Celery Worker for async background jobs:

python -m celery -A app.worker.celery_app worker --loglevel=info -P solo

3. Frontend (React/Vite)

Start the frontend user interface:

cd "1. Source_code/frontend"

# Install Node dependencies
npm install

# Start Vite development server
npm run dev

Note: The frontend web app will open at http://localhost:5173.

---

🎥 Video Demonstrations (Proof of Work)

To evaluate Prism AI's capabilities, we have provided comprehensive video evidence of the system operating under different conditions.

🔗 View All Demo Videos on Google Drive

1. Offline Working Proof

Demonstrates the platform running 100% locally with the network adapter disabled (Wi-Fi/Ethernet off). Proves that LLM inference, embedding generation, and guardrails function entirely on the host hardware without cloud APIs.

2. Online Working (Live Demonstration)

Showcases the full workflow in a standard environment. Highlights the speed and fluidity of the React/Vite UI, generating multiple formats (e.g., PDF to Twitter Thread & Video Script) from a single prompt, and instant preview functionality.

3. Database and Backend Working

Provides a technical deep-dive into the 3-Database Architecture and pipeline flow. Demonstrates the Docker containers running (PostgreSQL, Qdrant, MinIO, Redis) and FastAPI logs processing Celery background tasks (RAG, Guardrails) in real-time.

---

💻 Usage & Workflows (Demo Script)

If you are evaluating or presenting PRISM AI, we recommend these core workflows:

Note: For Usage & Workflows (Demo Script), please refer to Deliverables _1 and _2 in the Deliverables Folder. You can find the sample data required for these workflows in the Sample Data Folder.

• 🎬 TEXT → VIDEO SCRIPT

  • Topic: "Cybersecurity's Need in Today's World"
  • Result: Generates a 2-minute educational video script, voiceover narration, timing breakdowns, and visual scene cues.

• 🎬 PDF/DOCX → TWITTER/X THREAD

  • Topic: Upload a Python course book.
  • Result: Extracts the "Functions" chapter into a progressive 5–7 post viral thread with technical concepts and code snippets.

• 🎬 PDF/DOCX → CYBERSECURITY ADVISORY

  • Topic: Upload Cybersecurity_Threats.pdf.
  • Result: Generates an 8-section formal advisory (Threat Overview, Affected Users, Attack Mechanics, Prevention, etc).

All workflows support configuring Tone (Professional/Technical), Audience, Language, and Custom Few-Shot Templates.

---

📂 Project Directory Structure

JudeForce-prism ai/
├── Readme.md                     # Main Project Documentation & Quickstart
├── 1. Source_code/
│   ├── docker-compose.yml        # Docker orchestration (Postgres, Qdrant, MinIO, Redis)
│   ├── backend/                  # FastAPI backend, RAG pipeline, guardrails, models
│   │   ├── main.py               # Entrypoint for FastAPI server
│   │   ├── test_phase5_fast.py   # Automated backend pipeline test suite
│   │   ├── app/                  # Routes, core logic, tasks, database schemas
│   │   └── requirements.txt      # Python dependencies
│   └── frontend/                 # Vite + React frontend application
│       ├── src/                  # React components, stores, hooks, custom CSS
│       └── package.json          # Frontend Node dependencies
├── 2. Architecture_Doc/          # Architectural specs & technical design docs
├── 3. Demo_Video/                # Demo video recording assets
├── 4. Deliverables for Evaluation/ # Evaluation artifacts & submission files
├── 5. Technical_PPT/             # Technical presentation decks
├── 6.Output_Screenshots/         # Application workflow screenshots
├── 7.basemodel_finetuning.ipynb - Colab.pdf # Documentation for base model finetuning
└── 8.base model _ dataset_adapters/ # Model adapters and dataset resources

---

🔒 Security & Privacy Statement

Prism AI is engineered with a strict Privacy-First Mandate.

All LLM inference, embedding generation, vector search, database operations, and document parsing run 100% locally on host hardware. No confidential prompts, documents, or metadata are ever transmitted to third-party cloud APIs unless explicitly enabled by the user. Outputs are protected by SHA-256 tamper-evident integrity hashes to ensure compliance and prevent malicious tampering.