GenAI Engineer | Building Production-Ready AI Systems That Scale

GenAI Engineer at Capgemini, focused on building enterprise-grade AI systems that move from experimentation to real, scalable products.

I work at the intersection of LLMs, product thinking, and cloud architecture, helping teams go beyond proofs of concept into reliable, secure, and observable GenAI solutions aligned with real business constraints.

My expertise spans end-to-end AI architecture, Agentic Systems, and LLMOps, with hands-on experience designing, deploying, and maintaining production-ready AI workflows. I frequently bridge technical teams and business stakeholders to ensure feasibility, robustness, and long-term value — not demos for demos' sake.

Certified as a Microsoft Azure AI Engineer, with 9 additional AI-focused certifications accross different clouds, I emphasize applied AI, observability, governance, and measurable outcomes in regulated and enterprise environments.

📊 Selected Impact

| Metric | Result | |--------|--------| | Automation efficiency for business processes | +40% | | Data classification and extraction workflows | +50% faster | | Conversational AI accuracy | +35% | | Solution relevance using RAG and vector databases | +30% |

🛠️ Technical Stack

Languages & Backend: Python, TypeScript, React, FastAPI

LLMs & Frameworks: LangChain, LangGraph, CrewAI, AutoGen, LlamaIndex

Cloud & Infrastructure: Microsoft Foundry, AKS (Azure Kubernetes Services) AWS Docker, Azure Devops, Redis (queues & caching)

Data & Retrieval: MongoDB, PostgreSQL, Vector Databases (Pinecone, Qdrant, Milvus, ChromaDB)

Observability & Evaluation: LangSmith, Langfuse, tracing, prompt evaluation, monitoring, custom metrics

Automation: n8n, low-code tools

🎓 Certifications

• Microsoft Azure AI Engineer Associate (AI-102)
• Databricks Generative AI Engineer Associate
• Google Generative AI Leader (GCP)
• AWS Certified AI Praticcioner
• Salesforce AI Specialist
• Salesforce AI Associate
• Oracle AI Foundations
• Microsoft Azure AI Fundamentals (AI-900)

🏆 Featured Projects (In this Repository)

A selection of key projects demonstrating end-to-end AI solution delivery. Click on the titles to explore the documentation for each system. Code and client infos are anonymized per NDA compliance

⚡ High-Performance Customer Support Crew — Multi-Agent AI Pipeline

Problem: Customer support centers suffer from massive ticket volume and high operational costs. Simple inquiries clog high-touch queues, delaying responses to critical issues and driving up API token costs.

Solution: Architected a multi-agent ticket triaging and resolution pipeline using CrewAI and Anthropic Claude. The system features a lightning-fast Claude Haiku classifier that dynamically routes simple inquiries via an Express path (under 5 seconds) and reserves the full multi-agent sequential workflow (Sonnet & Opus) for complex inquiries. Implements native Anthropic Prompt Caching for 80% cost reduction, real-time log streaming using FastAPI SSE, an LLM-powered Semantic Cache, a local SQLite DB tuned in WAL mode, secure JWT auth, and a responsive glassmorphic operations dashboard.

Stack: Python, FastAPI, CrewAI, Anthropic Claude, SQLite (WAL), Semantic Caching, HTML5/CSS3 (Glassmorphism), JWT, SSE (Server-Sent Events), Langfuse.

Impact: Achieves an 85% response time reduction for routine support tickets, up to 80% savings in LLM token costs, and total data safety with custom regex-based PII anonymization and secure Human-in-the-Loop auditing.

🚀 Strategic Sourcing Intelligence Engine — MCP Server

Problem: Procurement teams spend hours manually auditing commercial proposals and spreadsheets. Static analysis tools fail to handle unstructured financial data and complex tax logic across various file formats.

Solution: Developed a high-performance MCP (Model Context Protocol) Server that allows LLMs to act as specialized procurement agents. Implemented semantic header mapping for template-less data discovery and a multi-tier delta calculation engine for 100% audit accuracy. Built a "Stateless Restore" pattern to maintain complex analysis state across tool calls.

Impact: Transforms hours of manual spreadsheet auditing into seconds of AI-driven analysis. Provides automated premium dashboards and legal-ready governance PDFs.

Stack: Python, MCP (Model Context Protocol), Pandas, NumPy, AsyncIO, XlsxWriter, ReportLab.

🌌 Violeta: Enterprise-Grade Multi-Agent Orchestration Ecosystem

Problem: Organizations face "Agent Sprawl"—where agent logic, prompts, and model configurations are hardcoded, fragmented, and lack centralized governance or observability as AI adoption scales.

Solution: Architected a modular, scalable framework that decouples agent configuration from execution logic, providing a robust Control Plane (Registry) and Data Plane (SDK) that standardizes the development lifecycle. Features a Centralized Registry for versioned configuration, a shared Security Layer (ms-auth-api) for RBAC authentication, and a developer SDK & CLI.

Impact: Empowered 150+ developers across multi-disciplinary business units (including HR, Finance, and Fraud Detection) to build, secure, and monitor agents; reduced production troubleshooting time by 60%; standardized authentication; and lowered operational LLM API costs via rate limiting and caching.

Stack: Python, TypeScript, LangChain, FastAPI, MongoDB, Kubernetes (AKS), Helm, Docker, Nginx, Azure DevOps, LiteLLM, Foundry, Langfuse.

🔧 AI-Powered Pull Request Automation — Autonomous Code Fixing System

Problem: Development teams spend significant time on repetitive bug fixes and manual PR creation. Error-to-fix cycles are slow, and there's no systematic way to ensure proposed changes are safe and minimal before human review.

Solution: Built an end-to-end automation system that transforms error reports into validated pull requests. Uses Claude Opus for error analysis and constraint planning, Claude Sonnet for minimal diff generation. Features a 6-layer validation system (status, confidence, diff existence, language consistency, change limits, line verification).

Impact: Automated the entire error-to-PR pipeline while keeping humans in the loop. PRs include AI justification and confidence levels.

Stack: n8n, Claude models with model routing, GitHub API, Webhooks, CI/CD

🎯 Agentic Market Research Team — Autonomous AI Intelligence System

Problem: Companies struggle to understand their customers' psychological drivers. Managing multiple AI agents typically requires complex orchestration and lacks persistent learning.

Solution: Architected a Level 5 autonomous agent system with 6 specialized AI agents that decode customer psychology. Features persistent memory via Qdrant vector database and collaborative LangGraph workflows, accessible through natural Slack commands.

Impact: 3x higher conversion rates using extracted language patterns. Reduced market research time from weeks to hours with continuous autonomous improvement.

Stack: Python, LangChain, LangGraph, Anthropic Claude, Qdrant, Slack Bolt, FastAPI, Docker.

💡 What I Bring

I thrive in global environments where innovation meets execution, building AI systems designed to scale, perform, and deliver sustained ROI. My approach emphasizes:

• Production-first mindset: Solutions built for reliability, not just impressive demos
• Observability & governance: Tracing, monitoring, and compliance for enterprise environments
• Business alignment: Bridging technical teams and stakeholders for measurable outcomes
• End-to-end ownership: From architecture to deployment to maintenance

📫 Let's connect — I'm always interested in challenging GenAI problems at enterprise scale.