Skip to main content
AI

AI Solutions Assistant

Next-gen portfolio interaction using RAG and Action Protocol.

Personal Project2025
<0.5s
Latency
Avg. TTFT
-95%
Cost
vs GPT-4o
99.9%
Uptime
Serverless

Tech Stack

Next.jsTypeScriptGroq (Llama 3)RAGPrompt Engineering

Context: The New Era of Digital Interaction

In today's landscape, static interfaces are becoming obsolete. Users (and clients) expect immediacy and personalization. As a Solutions Engineer, I identified an opportunity: transform the personal portfolio from a 'digital brochure' into an active agent capable of selling, explaining, and navigating by itself. This isn't just a 'cool' feature; it's a case study on how Generative AI can be integrated into real business workflows without friction.

Security & Privacy: Enterprise-Ready

Is it secure? Yes. Security is the #1 priority in enterprise environments. This solution implements a Zero-Trust architecture on the frontend: 1. API Proxying (Edge Layer): Groq/OpenAI API keys never touch the client's browser. All requests pass through an intermediate server (Next.js Edge Runtime) that securely injects credentials. 2. Data Privacy: The current implementation follows a 'Stateless Conversation' policy. We do not store persistent logs of user conversations in a database, ensuring that sensitive questions about strategy or business are not permanently recorded. 3. Input Sanitization: Protection against 'Prompt Injection' via strict Zod schema validation before processing any action command.

Technical Demo: Action Protocol

What makes this solution unique is its ability to 'touch' the interface. Unlike a standard chatbot that only returns text, this agent emits Structured Intents (JSON). The frontend listens for these intents and executes state changes (Navigation, Highlighting, Scroll).
Demo: The agent interpreting natural language ("show me x") and executing autonomous navigation.
Demo: The agent interpreting natural language ("show me x") and executing autonomous navigation.

Economic Viability & Scalability

Cost is decisive. Implementing GPT-4 for a public customer service can be ruinous. This architecture uses Llama 3 on Groq, reducing inference costs by 95% compared to massive proprietary models, without sacrificing intelligence for navigation tasks. Latency: <300ms (Time-to-First-Token). Real-time feel. Scalability: Being Serverless, the infrastructure scales to zero when not in use (0€ cost) and can absorb traffic spikes from thousands of concurrent users without additional configuration.

Enterprise Applicability

This same architecture is directly translatable to internal Fortune 500 use cases: Employee Onboarding: A bot that not only explains how to request vacation but opens the HR form for you. Technical Documentation Search: RAG (Retrieval Augmented Generation) connected to Confluence/Jira to answer engineering questions with citations to private documents. • Smart Dashboards: "Show me Q3 sales" -> The bot navigates and filters the Tableau/PowerBI dashboard automatically. It is a modular, secure, and model-agnostic solution (plug-and-play with Azure OpenAI or Bedrock).

System Architecture

Decoupled design for maximum maintainability:
Secure Flow: Next.js Edge -> LLM Inference -> Client Action Handler
Secure Flow: Next.js Edge -> LLM Inference -> Client Action Handler
Next.js
TypeScript
Groq (Llama 3)
RAG