Portfolio | Custom Software & AI Engineering

Toptal Network MemberTop 3% of talent worldwide

My Latest OS Projects

RedAmon

Unmask the hidden before the world does.

An AI-powered agentic red team framework that automates offensive security operations, from reconnaissance to exploitation to post-exploitation, with zero human intervention. Fully containerized, it chains six-phase reconnaissance, LangGraph-based autonomous agents, and Neo4j-powered attack surface mapping into an end-to-end offensive security pipeline.

Project repository

Watch the video Visit redamon.org

PATHBREAK

See the attack path. Then break it.

Agentic Cloud Security

Pathbreak is an agentic cloud security platform that autonomously discovers and validates the real attack paths a threat actor would exploit across your cloud, then shows you exactly how to break them. It reasons across identities, misconfigurations, and network exposure to chain individual weaknesses into full, exploitable attack paths and prioritize what actually matters.

Visit pathbreak.io

My Latest Research Paper

arXiv2610.14216cs.SE · cs.AIOctober 2026·46 pages, 22 figures, 8 tables

Cortex: Measurement-Gated Evolution and Pruning of Coding-Agent Skills and Rules

Samuele Giampieri

Coding agents such as Claude Code follow text files that developers write for them: a project file, skills and rules. Together they form the agent's harness, and it grows every time a developer corrects the agent, yet one passing run proves little. Cortex tests harness changes instead of trusting them: it turns the developer's corrections into replayable tasks, writes one skill or rule at a time, and keeps it only when repeated runs with and without it show a net gain and no task made noticeably worse. Existing text faces the same test before it is pruned. The tool, the lab and the data are open source.

Read on arXiv Download PDF

Code and data

MY OS PROJECTS

In this section you'll find my open source projects that I share with the community. For each project, I've created a YouTube tutorial explaining the core ideas and development logic in detail, guiding developers through the integration of the solutions I present. These projects are experimental and MVP-style, but they are then integrated into real production systems using the same concepts, expanded and refined for enterprise deployment. I firmly believe that if we've achieved incredible milestones in IT today, it's thanks to open source, knowledge sharing, and the developer community. This is my contribution to a better world built on sharing our work.

Redamon: AI-Powered Agentic Red Team Framework for Automated Offensive Security

Watch the video

Project repository

Real-Time Cyber Attack Detection System with ML & AWS (Full MLOps Pipeline)

A production-grade, real-time cyber attack detection system deployed on AWS with automation. Features autoencoder-based anomaly detection, incremental SGD classifier for streaming data, dual training pipelines (batch & streaming), AWS CloudFormation infrastructure, CI/CD, Blue/Green deployments, and attack simulation for DoS, exploits, backdoors & more.

Watch the video

Project repository

Vision RAG: Multimodal Document Search with Text & Image Embeddings

The future of document search is here! Vision RAG fuses text and image processing into one powerful RAG workflow. Features dual embedding strategies (OpenAI + Cohere), PostgreSQL with pgvector for semantic search, Docker-first deployment, unified ingestion layer with OCR support, and multimodal answers combining text & images using GPT or Gemini.

Watch the video

Project repository

Cortex: Measurement-Gated Evolution and Pruning of Coding-Agent Skills and Rules

Cortex tests changes to a coding agent's harness (the project file, skills and rules Claude Code follows) instead of trusting them. It turns the developer's corrections into replayable tasks, writes one skill or rule at a time, and keeps it only when repeated runs with and without it show a net gain and no task made noticeably worse. Existing text faces the same test before it is pruned. Published on arXiv.

Project repository

Paper on arXiv

Build a Large Language Model from Scratch with Transformers & Keras

Demystify LLMs! A visual deep dive into Transformers: self-attention, embeddings, positional encoding, multi-head attention, and feed-forward networks explained intuitively. Then build a text generation model from scratch in Keras tokenization, dataset preparation, training with real metrics, and output analysis. No black boxes, just real code and clear math.

Watch the video

Project repository

Hybrid Multi-Vector Knowledge Graph RAG with Advanced Graph Traversal

Combine vector embeddings and graph databases for intelligent retrieval! Features Neo4j integration, multi-vector embeddings, and advanced traversal algorithms: Context-to-Cypher, K-hop BFS, Depth-Limited DFS, Uniform Cost Search, A* Heuristic, Beam Search all with optional LLM predicate filtering. Includes entity extraction with LangChain and visualization techniques.

Watch the video

Project repository

AI Agent Framework Benchmark: LangChain, OpenAI SDK, CrewAI & More

A comprehensive benchmark of top AI agent frameworks (LangChain, OpenAI Agent SDK, CrewAI, LlamaIndex, AutoGen, Semantic Kernel) vs pure Python. Covers multi-agent orchestration, state management, tool integration, MCP interoperability, code execution pipelines, and guardrails. Each framework scored on abstraction, readability, setup complexity, and flexibility.

Watch the video

Project repository

Semantic Text2SQL: RAG + Text-to-SQL Combined for Smart Database Queries

Why choose between RAG and Text-to-SQL when you can have both? This system combines Traditional SQL for precision, Vector Embeddings for semantic understanding, and Fuzzy Matching for typo tolerance. Features multi-language support, self-healing queries, and SQL injection protection.

Watch the video

Project repository

Transformer-Based Time Series Forecasting: Multi-Head Attention for Stock Price Prediction

A Transformer-based time series forecasting model for predicting stock-price signals using TensorFlow and Keras. Features multi-head self-attention, stacked Transformer blocks with residual connections, sliding-window datasets, positional encodings, TensorBoard monitoring, and visualization tools for capturing long-range dependencies in financial time series.

Watch the video

Project repository

Model Context Protocol (MCP) Server: FastAPI & WebSocket Implementation

A Model Context Protocol server built from scratch using FastAPI, WebSockets, and async/await in Python to understand low-level MCP logic. Features dual-mode operation (HTTP REST and WebSocket), server capabilities (tools, prompts, resources), API-key authentication, and a client agent that converts natural language to SQL queries.

Watch the video

Project repository

SurfAi - AI-Powered Web Automation Agent with Playwright

SurfAi is an intelligent and lightweight web automation engine that leverages AI to interpret natural language instructions and automate web interactions using Playwright. It combines LLM capabilities with browser automation for complex task execution. For exmample it can book flights, buy things online, or login and create automatic posts on LinkedIn.

Watch the video

Project repository

Optimize LLM Inference: KV Caching & Causal Masking from Scratch

A Transformer-based LLM implementation featuring Key-Value (KV) Caching and Causal Masking for production-grade text generation. This project demonstrates GPT-style architecture in Python/Keras/TensorFlow with efficient O(n) decoding, dynamic cache updates, causal attention masks, and performance benchmarks showing dramatic speedup compared to vanilla implementations.

Watch the video

Project repository

AutoCodeAgent 2.0: Intelligent Task Decomposer with RAG ingestion/retrieval as tools

What happens if you add RAG ingestion/retrieval as tools to AutoCodeAgent? We have AutoCodeAgent 2.0! An intelligent system that can automate ingestion and retrieval in the db to enhance the agent's performance and use them as a subtask tools to solve complex tasks.

Watch the video

Project repository

Rushes: Turn Any Web App into a Narrated, Captioned, Verified Demo Video

Point your coding agent at your web app and get a demo video back. You describe the demo, the agent writes a storyboard, and Rushes drives your real UI in a real browser with Playwright, narrates each scene with ElevenLabs, and cuts it together with ffmpeg into a video with captions and chapters. Then it verifies the result: no unresolved step, no dead air, no secrets on screen. Works with Django, Rails, Next.js or a static site.

Project repository

Scruple: Certified, Abstaining Qualitative Coding with a Publishable Kappa

Qualitative coding that codes what it is sure of, refuses the rest, and proves both. Where a general-purpose LLM lands around κ = 0.57 against the κ ≥ 0.70 publication bar, Scruple uses calibrated probabilities to code only what it can defend, reaching κ = 0.99 after a person hand-codes 16% of the corpus instead of 95%. Ships the coded dataset, the abstentions, and a validation report with full provenance.

Project repository

Fine-Tuning Gemma 3 1B-IT for Math Reasoning: LoRA, GRPO & Unsloth

Fine-tune Google's Gemma 3 1B-IT model for advanced math reasoning using LoRA (Low-Rank Adaptation) and GRPO (Generative Response Policy Optimization). Features Unsloth optimizations for 4-bit quantization, custom reward functions for numeric accuracy and formatting, chain-of-thought structuring with XML tags, and training on consumer-grade hardware. Includes full code walkthrough for dataset preparation, reward shaping, and model deployment.

Watch the video

Project repository

AutoCodeAgent: Intelligent Task Decomposer and Code Generator

AutoCodeAgent is an advanced AI-powered system that automatically breaks down complex tasks into manageable subtasks, generates and executes Python code for each, and iteratively refines the process. Leveraging cutting-edge AI techniques, it orchestrates the use of various tools and libraries to solve problems efficiently and effectively.

Watch the video

Project repository

Fine-Tuning LLM for Tool Selection: Supervised Fine-Tuning with Hugging Face

A workflow for adapting compact language models to select software tools. Features synthetic data generation for 10,000 query-tool pairs, tokenizer extension with control tokens, TRL's SFTTrainer configuration, 4-bit quantized training on consumer hardware, and continuous evaluation. Uses SmolLM2-135M from Hugging Face.

Watch the video

Project repository

Solve complex task with an AI Interactive Multiagent

An intelligent system designed to orchestrate and manage a chain of autonomous agents to solve very complex tasks. By decomposing intricate tasks into simpler, manageable sub-tasks, Interactive Multiagent ensures efficient and effective task execution through specialized agents. It's a sort of chain of thought where agents are interacting with each other and the user to solve the task.

Watch the video

Project repository

AI Blog Generator: Automate Trends Blogging with LangChain and GPT-4

The AI Blog Generator is a Dockerized system that leverages LangChain, OpenAI's GPT-4, and Hugging Face APIs to create engaging blog articles on trending topics. The workflow includes automated trend analysis, news extraction, detailed article generation, and custom image creation.

Project repository

Designing an AI Agent with a Graph-Based State Machine Using PyTransitions

Sophisticated AI agents depend on workflows managed by state machines. While many open-source libraries exist (e.g., LangGraph), PyTransitions stands out due to its broad functionalities and robust extensions like HierarchicalMachine, GraphMachine, and AsyncMachine, making it an excellent choice.

Watch the video

Project repository

AI-Powered Web App: Integrating Generative AI with Next.js & AWS

This advanced web app leverages cutting-edge Generative AI, featuring an AI super-agent for web search, data processing, and LLMs. It includes custom AI training, AWS S3 file handling, and real-time communication via Socket.io. Built on Amazon Bedrock with AWS Lambda and OpenSearch integrations, the app delivers smart, scalable, and high-performance solutions.

Watch the video

Project repository

Scraping with Advanced Anti-Detection Techniques

The system adopts advanced anti-detection techniques: undetected_chromedriver for stealth browsing, proxy and user-agent rotation, simulation of human interactions, spoofing of navigator properties, and overriding JavaScript methods to avoid fingerprints. It operates in headless mode with pyvirtualdisplay, creates temporary profiles, and updates libraries against anti-bot techniques.

Project repository

Project repository 2

Text-to-SQL Query System with LangChain

This interactive system translates natural language questions into SQL queries using Open AI and LangChain, executes them against a SQLite database, and provides detailed answers. It features database integration, customizable prompts, and an interactive CLI for real-time querying.

Project repository

MNIST Handwritten Digit Recognition: MLP with Keras & TensorBoard

A tutorial for training a Multi-Layer Perceptron (MLP) to recognize handwritten digits from MNIST using Keras. Features data preprocessing, Keras Functional API with BatchNormalization and Dropout, TensorBoard callbacks for monitoring, Visualkeras visualization, metrics (confusion matrix, precision, recall, F1-score), and model saving/loading. Includes sample predictions and result reporting.

Watch the video

Project repository

Collaborations

Freelance logo I am an AI Platform Architect & Full-Stack Lead with over 15 years of experience building production-grade software systems. My core expertise lies at the intersection of software engineering and artificial intelligence, designing scalable, robust applications that seamlessly integrate AI capabilities for real-world enterprise environments.

Certifications: I am multi-certified in AWS for cloud expertise and IBM certified for AI engineering, bringing proven credentials to every project.

Toptal Network Member: I am a proud member of the Toptal network, having passed its notoriously selective, multi-stage screening process that accepts only the top 3% of talent worldwide. This vetting covers advanced technical assessments, live problem-solving sessions, and real-world test projects, independently validating both my engineering excellence and my professionalism with demanding international clients.

AI & Machine Learning Expertise: I architect and deploy end-to-end ML solutions covering supervised, unsupervised, and reinforcement learning. My deep learning experience spans CNNs, RNNs, LSTMs, and Transformer architectures, with hands-on expertise in model training, validation, hyperparameter tuning, transfer learning, and fine-tuning for production deployment.

AI Specializations: I design sophisticated NLP systems (text processing, sentiment analysis, chatbots, language models, LLM fine-tuning), Computer Vision solutions (image classification, object detection, segmentation, image generation), hybrid RAG architectures with vector and graph databases (Pinecone, Weaviate, Chroma, Neo4j), and AI Agent systems using LangChain/LangGraph for autonomous, intelligent multi-agent workflows.

Cloud & DevOps: I specialize in deploying AI at scale on AWS (SageMaker, Bedrock, Lambda, EC2, S3, RDS). I implement robust DevOps practices including containerization (Docker, Kubernetes), automated CI/CD pipelines, production model serving, monitoring, and experiment tracking with MLflow and Weights & Biases. I ensure scalability, security, and reliability in cloud-native solutions.

Full-Stack Development: I have mastery in Python, PHP, JavaScript, Node.js, SQL, TypeScript, TensorFlow, and Keras. Beyond AI, I build fast, maintainable applications using modern frameworks like React, Next.js, Django, Flask, and Express.js. I design scalable APIs and integrate AI functionalities into production systems with proper software engineering practices.

Tools & Frameworks: My toolkit includes Hugging Face Transformers, LangChain/LangGraph for AI applications, vector databases for RAG systems, and comprehensive ML platforms for training and deployment. I emphasize clean architecture, version control, testing, and maintainable code across all projects.

I am passionate about solving complex problems with AI-driven solutions that actually work in production. Constantly exploring cutting-edge technologies, I focus on delivering enterprise-grade systems that are secure, scalable, and built for long-term success.

I offer my services as an AI Platform Architect & Full-Stack Lead to collaborate with companies of all sizes. Whether as a full-time partner or on a project basis as a freelancer (with VAT number), I am committed to delivering robust, production-ready solutions.

If you are interested in collaboration or desire more information, please do not hesitate to contact me. I am here to help you build intelligent, scalable systems that drive real business value.

Stacks e Frameworks

AI, LLM & Agents