Skip to content
View EsraaKamel11's full-sized avatar

Block or report EsraaKamel11

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
EsraaKamel11/README.md

👋 Hi, I'm Esraa Kamel

Senior Agentic AI Engineer | LLM Orchestration & Multi-Agent Systems Specialist

Banner

🚀 About Me

I am a Senior Agentic AI Engineer specializing in the orchestration of autonomous, goal-oriented AI systems. My work focuses on moving beyond simple chat interfaces to building production-grade agentic workflows that possess reasoning, planning, and self-healing capabilities. I architect multi-agent systems that solve complex, non-linear problems at scale.

  • 🔭 Currently Architecting: A distributed multi-agent swarm for automated software engineering.
  • 🧠 Core Expertise: LLM Reasoning (CoT/ToT), Agentic RAG, Tool-Use Optimization, and Multi-Agent Orchestration.
  • 🌱 Deep Diving: Exploring the intersection of Small Language Models (SLMs) and edge-based agentic reasoning.
  • 👯 Open Source: Active contributor to aden-hive/hive (Agentic Framework Internals).
  • 💬 Consulting: Helping enterprises transition from "Chatbots" to "Agentic Workforces".

🏗️ Engineering Philosophy: Designing for Failure

I build production agentic AI systems where failure is handled, not hoped away. While most systems are designed for the "happy path," I design for the rate limits at step 3, the approvals that arrive 4 hours later, and the platforms that redesign their UI overnight.

My 6 Hard Constraints for Production Agents:

  1. Single Responsibility: Every agent answers exactly one question to isolate failure modes.
  2. Deterministic Routing: Routing is pure Python, not LLM-driven, for testability and version control.
  3. Mandatory Verification: Post-action verification is a hard checkpoint before state advances.
  4. Durable HITL: Human-in-the-loop is a named graph node that survives process restarts and deployments.
  5. Architectural Boundaries: Action limits are compile-time constants, not prompt-level suggestions.
  6. Hybrid Retrieval: Dense + sparse retrieval with cross-encoder re-ranking for real-world query variance.

🛠️ Signature Architectural Patterns

These patterns are the foundation of every production system I build, ensuring reliability, security, and resilience in non-deterministic environments.

Pattern Description Justification
PAV (Perception-Action-Verify) A vision-guided loop where a PerceptionAgent analyzes screenshots to guide actions and verify outcomes. Eliminates reliance on brittle CSS selectors; ensures the agent "sees" the result of its work.
UNVERIFIABLE State A first-class state for actions that executed correctly but cannot be confirmed due to propagation delays. Prevents duplicate actions from retries and maintains system honesty during async processing.
API-First with Browser Fallback Automatic transition to browser automation when API endpoints fail or are unavailable. Ensures 100% task completion regardless of external API stability or coverage.
Capability Self-Calibration Real-time reliability tracking that auto-disables/re-enables tools based on health checks. Prevents cascading failures and optimizes routing toward the most stable execution path.
Audit Hash Chains Append-only logs where each row contains a SHA-256 hash of itself and the previous row. Provides tamper-evident proof of every agent action and decision for high-stakes environments.

🛠️ Tech Stack

Category Tools & Technologies
Agent Frameworks LangGraph CrewAI PydanticAI Semantic Kernel
LLMs & APIs GPT-4o Claude 3.5 Llama 3 Mistral
Vector DBs Pinecone Weaviate Milvus Chroma
Backend & Infra Python FastAPI Docker Kubernetes
Monitoring & Eval LangSmith Weights & Biases Arize Phoenix

🌟 Featured Projects

🛡️ Security & Adversarial AI

  • AI Product Red-Team Agent: LangGraph · AutoGen · Claude API · FastAPI · Redis. Lethal Trifecta injection · CVSS-scored vulnerability reports.
  • Automated Code Compliance & Security Agent: AWS Bedrock · Guardrails · GitHub Webhooks · Terraform. 100% policy adherence · 60% reduction in code review time.
  • Agentic Supply Chain Orchestrator: Azure AI Foundry · RAG · GPT-4o · 5 agents. BOM → PO with zero manual steps · 100% error detection.

🤖 Autonomous QA & Testing

  • Autonomous QA Agent Framework: LangGraph · ChromaDB · GitHub Actions · pytest · 8 agents. Full test lifecycle: PR diff → generation → self-healing → defect filing.
  • Self-Healing Selenium Framework: LangGraph · Selenium · PostgreSQL · 6 agents. DOM-aware repair · confidence-gated HITL · autonomous commit on high confidence.
  • Agentic Defect Intelligence Pipeline: LangGraph · Qdrant · Kafka · hybrid RAG · 6 agents. Closed-loop: production failures → pre-release test coverage feedback.
  • Multi-Agent User Simulation & Chaos Lab: LangGraph · LangChain · FastAPI · 4-layer architecture. Persona generation · chaos injection · behavioral anomaly detection.

🧪 Open Source Contributions

Python · Agentic Framework Internals

  • Impact: 9 core contributions across bugs and features.
  • Key PRs: #5923, #5855, #5918, #5760, #5805, #6214, #6542, #6555, #6605.

🎓 Credentials

  • 🎓 MSc in AI, Agentic AI & LLM Engineering — Woolf University (in progress)
  • 🎓 BSc Computer Science — Menoufia University
  • ☁️ AWS AI/ML Scholar 2025

📊 GitHub Stats


📫 Connect with Me

LinkedIn Gmail


💡 Quote of the Day

"The best way to predict the future is to build an agent that creates it." — AI Proverb

Visitor Count

Pinned Loading

  1. Autonomous-QA-Agent-Framework Autonomous-QA-Agent-Framework Public

    Production-grade multi-agent QA pipeline built with LangGraph that autonomously generates, executes, self-heals, and evaluates API test suites from PR diffs and OpenAPI specs. Features LLM-as-Judge…

    Python

  2. AI-Product-Red-Team-Agent AI-Product-Red-Team-Agent Public

    Autonomous multi-agent adversarial testing system for AI/LLM products. 35 attack vectors across 8 categories, 5-generation evolutionary mutation, AI-CVSS scoring, Lethal Trifecta indirect injection…

    Python

  3. Agentic-Defect-Intelligence-Pipeline Agentic-Defect-Intelligence-Pipeline Public

    Multi-agent AI pipeline for production defect intelligence. Ingests CI/CD, Jira, and Sentry streams, clusters failures via HDBSCAN-over-UMAP, scores file-level risk with a Bayesian-adaptive formula…

    Python

  4. Simulation-Chaos-Lab Simulation-Chaos-Lab Public

    Multi-Agent User Simulation & Chaos Lab — A 12-agent LangGraph pipeline that generates diverse user personas, executes multi-step journeys with vulnerability-guided chaos injection, detects anomali…

    Python

  5. Self-Healing-Selenium-Framework Self-Healing-Selenium-Framework Public

    AI-powered self-healing Selenium framework that autonomously detects broken test selectors, generates ranked repairs via Claude Sonnet + LangGraph, validates fixes through multi-layer protocols, an…

    Python

  6. UDA-Hub-Multi-Agent-Customer-Support-System UDA-Hub-Multi-Agent-Customer-Support-System Public

    Supervisor-based LangGraph multi-agent customer support system that Features RAG, persistent memory, tool integration, and confidence-based routing for intelligent ticket resolution.

    Python