I am a Full Stack & AI Developer specializing in AI Agent Architecture, Multi-Agent Autonomous Systems, and Production-Grade SaaS Platforms. With a strong background in Mathematical Sciences and Software Engineering, I bridge the gap between frontier LLM research and robust, enterprise-ready software.

My core expertise centers around architecting self-hosted AI agent frameworks (OpenClaw), novel protocol patterns (such as Semantic Tool Contracts via Model Context Protocol / MCP), multi-model chat applications handling 500+ frontier models, and secure offline enterprise RAG pipelines.

• 🤖 AI Agent Systems & Multi-Agent Swarms: Architecting multi-agent orchestration engines on OpenClaw, intent routing, tool-using autonomous agents, and zero-selector web scraping pipelines (Firecrawl API).
• 🧩 Semantic Tool Contracts (MCP): Pioneered tool architecture patterns where MCP microservices return explicit semantic execution states (NO_DATA_FOUND, SUCCESS_WITH_DATA, ERROR) and active dialogue guidance.
• 🏗️ Full-Stack SaaS Architecture: Building high-throughput, multi-tenant applications with Next.js 16, React, Node.js, NestJS, TypeScript, PostgreSQL, and Prisma ORM.
• 📐 Hybrid Text-to-SQL (LEGO Assembly): Engineered modular Text-to-SQL pipelines combining pre-verified schema joins with flexible LLM filter clauses to solve the reliability-flexibility paradox.
• 🛡️ Enterprise Offline AI & Fine-Tuning: Designing zero-external-dependency intranet RAG engines using Ollama, Qdrant VectorDB, and fine-tuned domain models (CamemBERT, Detectron2 OCR).

Live Demo: qrox.vercel.app | Tech: Next.js 16, Puter.js AI API, PostgreSQL, Prisma, TailwindCSS, Shadcn UI

• Perplexity-inspired chat UI enabling interaction with 500+ frontier AI models (OpenAI, Claude, Gemini, Mistral, xAI) with zero API key management via innovative user-pays architecture.
• Dual-Authentication Engine: JWT + Puter SDK guest mode achieving <10s session initialization with automatic fallback and race-condition guards.
• Real-Time Streaming & Vision: Markdown rendering, syntax highlighting, multi-image vision support (10 concurrent uploads), and native AI image generation via function calling.
• High-Performance Persistence: Atomic conversation persistence with balance guards achieving 99.9% data integrity and <200ms database latency on Vercel/PostgreSQL.

Stack: OpenClaw, Docker, Firecrawl API, MySQL, OpenRouter (owl-alpha), Python/uv, Linux

• Self-Hosted AI Gateway: Deployed a 2-tier agent architecture on OpenClaw proxied to OpenRouter LLMs, Firecrawl web intelligence, and MySQL persistence.
• Autonomous RE-Scraper Subagent: Built zero-selector scraping autonomy that analyzes page structures, handles hidden APIs/PDFs/pagination, extracts 33 normalized fields per real-estate unit, and executes delta scraping with UPSERT deduplication.
• Production Performance: Extracted 262 rental units across 20 floors with 100% data integrity, automated unit conversion ($\text{sq ft} \rightarrow \text{m}^2$), and integrated OpenTelemetry diagnostics.

Confidential Client, Rabat | Stack: Python 3.11+, TypeScript, FastMCP, FastAPI, Ollama, Sentence Transformers, React, SQL Server, Docker

• Autonomous Natural-Language-to-SQL: Transformed personnel database queries into an autonomous French conversational AI interface.
• Semantic Tool Contract: Designed novel MCP tool states providing explicit dialogue guidance and next-action recommendations.
• Sentence Transformers & Self-Healing Validator: 3-tier entity matching (paraphrase-multilingual-MiniLM-L12-v2), Rust-compiler-inspired SQL auto-correction, and Differential Diagnosis menus for query ambiguity.

Status: Production-Ready MVP | Stack: Next.js, Node.js, PostgreSQL, Prisma, n8n, Chatwoot, Google Gemini API, FB Graph API, Docker

• Vertical E-Commerce AI SaaS: Transforming Moroccan merchants' Facebook Messenger inquiries into 24/7 AI sales agents with category-specific personalities for Fashion, Beauty, and Electronics.
• Ultra-Fast Performance: Multi-tenant architecture supporting 1,000+ concurrent stores with sub-50ms cached PostgreSQL queries (10–100x speed improvement over external APIs).
• Multi-Lingual Automation: Darija/French/English auto-detection, CSV/XLSX product import with auto-column mapping, and encrypted OAuth Facebook integration.

Confidential Client, Rabat | Stack: Ollama, Open WebUI, n8n, Docker, Qdrant VectorDB, Prompt Engineering

• Air-Gapped Privacy: Secure offline intranet deployment analyzing corporate documents, compliance, and legal support with zero external API dependencies using local LLMs (Qwen, DeVstral, GLM).
• Precision RAG Pipeline: Custom Qdrant VectorDB workflows achieving 95% accuracy in policy violation detection.

Stack: React, Express.js, Node.js, MySQL, Sequelize ORM, JWT, Tailwind CSS, Shadcn UI, Turborepo

• Scalable healthcare monorepo connecting patients with healthcare providers through real-time booking, doctor availability calendars, and role-based access control (RBAC).

Stack: Python, CamemBERT, FastAPI, Docker

• Internal ML pipeline fine-tuning CamemBERT for French-language sentiment analysis on employee performance evaluations, reducing review cycles by 60% with 92% human correlation.

Stack: Node.js, Express.js, MongoDB, React.js, FastAPI, Docker, Detectron2, JWT

• Automated Moroccan license plate recognition system utilizing a custom-trained Detectron2 OCR model (94% accuracy) serving Fast-API endpoints with 99.5% uptime.