My Projects

Claude-powered operational intelligence platform that connects to live engineering data over MCP and answers engineering-operations questions in natural language, producing a daily morning brief and live incident views.

Python backend server implementing the Model Context Protocol to host AI models and expose them to MCP-compliant agents and applications through a standardized interface.

Multi-tool Gradio application with MCP integration, bundling several tools behind a single interactive front end.

Security-first LLM fine-tuned for Agentic DevOps that translates natural language into safe, valid Bash commands while achieving a 100% adversarial attack blocking rate. Implements a Defense-in-Depth architecture with sanitized QLoRA fine-tuning and three-layer runtime guardrail system.

Layered security framework for LLMs and agents that combines deterministic rules, semantic classifiers, and policy enforcement to detect prompt injection, malicious intent, and unsafe reasoning before outputs reach users or systems.

Security-focused RAG application with built-in guardrails for safe document and codebase analysis with AI.

Enterprise-Grade Security: Advanced ensemble ML system with 9 trained models, grey-zone decision gating, and SHAP explainability. Features Streamlit web interface, CLI tools, and production-ready deployment achieving 95%+ recall for SOC environments.

AI-powered malware detection using LightGBM and static PE file analysis for enterprise endpoint security. Achieves 98%+ F1-score with millisecond prediction times.

Comprehensive malware detection using PE file metadata with advanced feature engineering, visualization, and model validation. Multi-method feature selection with PCA, t-SNE, and ensemble methods.

Integration Capstone: Combines supervised and unsupervised ML approaches with recall-first optimization for SOC environments. Synthesizes insights from LightGBM and PE metadata projects into a unified detection framework achieving 99.95% recall.

Proactive Malware Defense: Signature-free malware detection using unsupervised learning for Aegis AI Security's Sentinel Endpoint Agent. Achieves 93% AUC with Isolation Forest for real-time pre-execution threat analysis.

Malware Triage: Advanced ensemble-based PE file analysis combining Isolation Forest, K-Means, and One-Class SVM. Features interpretable anomaly scoring and confidence-based triage for SOC environments with rich visualization suite.

An intelligent security system with face detection and motion tracking, built with Python and OpenCV.

Guardrail-first agent security monitoring framework built around OpenClaw job-search telemetry, detecting unsafe agent behavior and preserving evidence for investigation.

Explainable malware classification using structural similarity. Classifies PE files as malware or benign by comparing them against known samples, surfacing the reasoning rather than returning an opaque risk score.

Security assessment of a financial-services chatbot and an e-commerce agent. Manual probing, automated scanning with Giskard, and adversarial simulation identified 16 distinct vulnerabilities across the OWASP Top 10 for LLMs.

Demo-safe SOC migration accelerator for organizations moving from outsourced SOC operations toward an internal model. A case-centric backend ingests SOAR-style payloads, normalizes evidence, applies deterministic guardrails, and produces structured triage and GRC output behind role-safe API views.

Data and training pipeline behind SecureCLI-Tuner: a security-aware ML workflow for translating natural language into Bash, built on Qwen2.5-Coder-7B-Instruct with zero-tolerance validation.

Portable governance pipeline for Claude Code - custom slash commands, validation hooks, and audit logging - built to demonstrate that an AI-augmented development workflow can be rolled out safely across a small team.

Offline-first rehearsal range for AI security workflows. Provides deterministic simulators and synthetic fixtures so a full loop - read the challenge, run the workflow, capture evidence, write the report - can be practiced without standing up lab infrastructure.

Step-by-step learning project on Generative Adversarial Networks and AI security - building image-generating models, then attacking and defending them to observe the failure modes directly.

Production-Grade Multi-Agent Security Evaluation Framework: Advanced evolution of Trust Bench featuring SecureEval guardrails, intelligent agent routing, consensus building, and operational excellence. Features specialized security, quality, and documentation agents with cross-agent collaboration, weighted scoring, confidence metrics, and comprehensive export capabilities. Includes structured logging, health probes, CI/CD pipeline, and 79% test coverage.

Secure multi-agent SOC triage with A2A guardrails and structured observability, designed as a v1 competition finalist with v2 upgrades planned for expanded SOC automation.

A LangGraph-based multi-agent workflow that inspects software repositories for security leaks, code quality gaps, and documentation health. Features a collaborative system of specialized agents working together to provide comprehensive repository analysis.

Proof-of-concept six-layer trust infrastructure for agent-native services, covering discovery, identity, trust signals, context disclosure, workflow validation, and liability evidence. Registers and verifies services rather than executing workflows itself.

Portable nine-agent software factory with security engineering built into the pipeline at both design time and implementation time. Tool-agnostic - installs into any project and produces auditable, security-reviewed code.

Multi-agent code-review pipeline with a deterministic merge gate, packaged as a versioned, parameterized, tested skill library. Built as a Claude Code Mastery course mini-project.

Sanitized public mirror of a private deployment, published for the Kaggle AI Agents Intensive capstone. Generates an AI assistant's global instructions and skill library for multiple runtimes from a single source of truth.

URL shortening service built spec-first, where every line of code answers to a documented requirement. Demonstrates spec-driven development and YAML prompt engineering; built as a Claude Code course assignment.

Expense-approval workflow on Google's Agent Development Kit where the routing gate is deterministic, not a model decision. A plain threshold comparison auto-approves small expenses and suspends larger ones on a RequestInput interrupt until a human approves or rejects, so the LLM never gates the money decision.

Retrieval-Augmented Generation system built to help healthcare professionals find medical information from trusted sources, using the Merck Manual as its primary knowledge base.

Sentiment analysis built on Hugging Face Transformers and PyTorch, demonstrating how to fine-tune a pre-trained transformer for a downstream classification task.

Neural network model for predicting bank customer churn, identifying accounts likely to leave so retention effort can be targeted.

XGBoost classifier predicting US visa application outcomes, used to surface the factors that most influence approval decisions.

Decision Tree classifier predicting personal loan acceptance, used to identify which existing bank customers are most likely to convert.

Exploratory analysis of food-delivery order data, surfacing customer ordering patterns and restaurant performance for a food aggregator.

XGBoost regressor applied to a continuous-target prediction task, covering feature preparation, training, tuning, and evaluation.

End-to-end Azure fundamentals — ARM-based VM deployment with tagging, governance policies, and Azure Monitor alerts. Demonstrates cloud architecture, cost optimization, and observability with Infrastructure-as-Code (IaC).

Dashboard that tracks Google Cloud BigQuery release notes, slicing daily updates into individual release items and filtering them by type. Flask server with a vanilla HTML/JS client.

Event-driven document ingestion on Google Cloud, provisioned end to end with Terraform. A GCS upload fires a storage notification onto Pub/Sub, an OIDC-authenticated push subscription invokes a Cloud Run service, Gemini extracts and classifies the document via Vertex AI, and the metadata lands in BigQuery behind a Streamlit dashboard.

A security gym for AI engineers to build muscle memory around sanitization, API management, and guardrail patterns using a Mastery Loop training engine.

Progressive Python training system organized as nine belts of ten kata each, using disciplined repetition, spaced review, and structured recall.

Self-updating personal study library for Claude Code that turns course material and product news into short reviewed lessons, kept organized and ready to study.

Timed exercise in systematically exploring, mapping, auditing, and documenting an unfamiliar codebase - the BackdoorBox research library - in under five hours.

Patient vital-signs REST API built as the hands-on teaching project for a Claude Code Mastery course week, used to demonstrate multi-session agent workflows against one codebase.

Market-data explorer that pulls price, RSI, short interest, and volume for a given ticker and ranks a watchlist of heavily shorted stocks by a composite score. Built as a data-analysis exercise, not investment advice.

Interactive terminal client for Google News: fetch headlines, search queries, filter by topic, and open articles in the browser with arrow-key selection. Reads the public RSS feeds, so it needs no API key.