Sentia.ai | Enterprise Artificial Intelligence (AI) Customer Relationship Management (CRM)

Intelligent Customer Experience & Fraud Detection for High-Volume E-Commerce

Sentia.ai is an enterprise-grade, fullstack Customer Relationship Management (CRM) microservice platform. Designed for the scale of modern e-commerce (targeting architectures similar to Flipkart, Amazon, and Oracle), Sentia leverages Large Language Models (LLMs), Retrieval-Augmented Generation (RAG), and Predictive Analytics to automate support workflows, detect fraud, and eliminate support latency.

📈 The Business Impact & Statistics

At massive transaction volumes, traditional human-in-the-loop support becomes an insurmountable bottleneck. Sentia.ai is engineered to aggressively optimize operational metrics based on industry-leading AI benchmarks:

• Drastic Cost Reduction: Traditional human-handled support tickets cost on average $6.00 to $25.00 per resolution. Sentia's Agentic Auto-Draft aims to reduce the cost of routine inquiries to $0.05 - $0.50, resulting in up to a 70% reduction in support OPEX.
• Microsecond Resolution Targets: While human queues have an average response latency of 5 to 30+ minutes, Sentia's LLM pipeline drafts contextually aware, RAG-backed responses in < 2 seconds, dramatically cutting Mean Time to Resolution (MTTR).
• Predictive Risk Mitigation: By proactively analyzing ticket metadata for fraud risk before the agent ever opens the ticket, Sentia prevents costly policy abuse and return fraud dynamically.

---

🧠 How Sentia Uses AI

Sentia isn't just a wrapper around an API; it runs a dedicated Deep Learning microservice to execute four core AI pipelines simultaneously whenever a ticket is submitted:

1. Intelligent Categorization (Zero-Shot Classification): The ML engine instantly reads the ticket description and categorizes it (e.g., Shipping, Refund, Technical, Product Inquiry). This eliminates manual triaging and ensures the ticket reaches the right department instantly.
2. Sentiment Analysis: Using NLP models, Sentia detects the emotional tone of the customer (e.g., Angry, Frustrated, Positive, Neutral). "Angry" or "Frustrated" tickets are visually flagged and bumped to the top of the queue for immediate VIP handling to prevent churn.
3. Predictive Fraud Risk Scoring: By cross-referencing ticket metadata against known anomaly patterns, the AI assigns a Fraud Risk probability (0-100%). High-risk tickets (e.g., serial refund abusers) are flagged with visual progress bars to warn agents before they process a payout.
4. Agentic Auto-Draft (RAG): Using Retrieval-Augmented Generation (RAG) and Qdrant Vector Search, Sentia searches a database of past successful resolutions and company policies. It then passes this context to an LLM to generate a complete, highly empathetic, and accurate draft response. Agents simply click "Approve & Send".

🔄 AI Pipeline Architecture (Data Flow)

graph LR
    A[React Client] -->|Submit Ticket| B(Node API)
    B -->|Persist| C[(MongoDB)]
    B -->|Webhook| D{FastAPI AI Server}
    
    subgraph ML Inference
    D -->|Category| E[Zero-Shot]
    D -->|Tone| F[Sentiment]
    D -->|Risk| G[Fraud Engine]
    D -->|Embed| H[Transformers]
    end
    
    H -->|Search| I[(Qdrant DB)]
    I -->|Context| J[RAG Prompt]
    J -->|Infer| K[LLM]
    K --> L[Auto-Draft]
    
    E --> M[AI Metadata]
    F --> M
    G --> M
    L --> M
    
    M -->|Insights| B
    B -->|UI Update| A

Loading

---

🏗 Enterprise Architecture & Deployment

Sentia.ai is structured as a decoupled, horizontally scalable 3-tier microservice architecture, designed for zero-downtime CI/CD deployments.

1. AI Inference Microservice (server/)

• Core: Python, FastAPI, Hugging Face, SentenceTransformers, Qdrant (Vector DB).
• Role & Compute Isolation: Machine learning inference (especially Transformer models) is exceptionally CPU/GPU intensive and synchronous by nature. By isolating this into a dedicated FastAPI microservice, we ensure that heavy LLM computations never block the Node.js event loop handling standard API traffic.
• Model Pipeline Details:
  • Zero-Shot Classification & Sentiment: Utilizes Hugging Face transformers (e.g., facebook/bart-large-mnli) to dynamically categorize text without hardcoded rules.
  • Vector Embeddings (RAG): Uses SentenceTransformers to convert customer descriptions and knowledge-base articles into high-dimensional vectors, stored and queried in Qdrant for lightning-fast semantic search.
• Production Deployment: 🚀 Hugging Face Spaces (Dockerized)
  • Deployed on dedicated GPU/CPU hardware through Hugging Face via a Docker container. This ensures optimized inference speeds and bypasses the severe cold-start latency and package size limits typical of AWS Lambda or Vercel when deploying heavy Python ML dependencies.

2. Core API Gateway & Business Logic (express_server/)

• Core: Node.js, Express, MongoDB, JWT.
• Role: The secure central nervous system. Handles stateless JWT authentication, Role-Based Access Control (RBAC), database persistence, and asynchronous orchestration of the Python ML microservice.
• Production Deployment: 🚀 Vercel (Serverless Functions)
  • Deployed as globally distributed Edge/Serverless functions on Vercel for infinite horizontal scaling and ultra-low latency data retrieval.

3. Client Interface (client/)

• Core: React, TypeScript, Vite, Framer Motion, Material UI.
• Role: A premium Single Page Application (SPA) offering a stunning dark-mode Admin Dashboard and a sleek Customer Portal.
• Production Deployment: 🚀 Vercel (Static Global CDN)
  • The Vite build is deployed to Vercel's global Edge Network, ensuring instant Time-To-Interactive (TTI) for customers worldwide.

---

💻 Tech Stack Matrix

Layer	Technologies Used	Production Environment
Frontend UI	React, TypeScript, Vite, Custom CSS, Material UI	Vercel Global Edge CDN
Backend API	Node.js, Express, MongoDB	Vercel Serverless Functions
AI / DL Server	Python, FastAPI, Hugging Face, Qdrant	Hugging Face Spaces (GPU/CPU)
Security	JWT, Bcrypt, RBAC	Environment Variables / Secrets

---

🚀 Local Development Setup

To run the complete platform locally, boot up all three microservices:

1. Python DL Microservice

cd server
python -m venv .venv
source .venv/bin/activate  # On Windows: .\.venv\Scripts\activate
pip install -r requirements.txt
uvicorn main:app --reload --port 8000

(Requires .env with HF_TOKEN, QDRANT_URL, and QDRANT_API_KEY)

2. Node.js Core Backend

cd express_server
npm install
npm run dev

(Requires .env with MONGO_URI, JWT_SECRET, and PYTHON_API_URL=http://localhost:8000)

3. React Frontend

cd client
npm install
npm run dev

(Runs on http://localhost:5173. Connects automatically to the Node backend.)

---

Built with a rigorous focus on clean code, decoupled scalability, and solving real-world enterprise bottlenecks.