Projects

Real-time LLM benchmarking platform comparing outputs from GPT-2, DistilGPT-2, BERT, and T5. Smart recommendation engine with domain-specific analytics cuts evaluation time by 70%.

RAG-powered chatbot that lets investors converse with complex financial reports in plain English. MongoDB Atlas vector search + Groq embeddings + LangChain orchestration.

Advanced computer vision platform combining multiple deep-learning models for real-time image analysis, object detection, and visual recognition in surveillance and inspection scenarios.

Comprehensive web application for tracking ecological data through AI. Surfaces extinction risk indicators and biodiversity trends from structured datasets. Built as part of the Electric Sheep Fellowship (Futurekind AI).

Voice-enabled application that captures spoken input and refines it into clearer, more structured prompts for LLMs via a preprocessing pipeline. Explored how prompt structure influences model reasoning quality and response reliability.

LLM pipeline that generates a title and then writes a full Medium-style article from a given topic. Demonstrates chained prompt engineering and structured generation.

Tableau dashboard aggregating usage, support, and billing data into a unified customer health score. Reduced churn by 23% and drove $2M+ expansion revenue at Contentstack.

Case study predicting parking validation outcomes using gradient boosting. Feature engineering on time-based and categorical variables with full EDA pipeline.

Protocol-level defense system for the Model Context Protocol (MCP). Zero-cost validation methodology achieved a 64.5% reduction in adversarial tool-use attacks.

Safe local replication of OpenAI's emergent misalignment fine-tuning experiment. Documented end-to-end methodology, observations, and safety considerations for the community.

Apart x Martian Hackathon submission. Mechanistic analysis of manipulation vulnerabilities in AI orchestration — identifies, ranks, and visualizes attack vectors against judge models.

Research prototype implementing circuit analysis, activation patching, and feature visualization using the OpenAI API. Exploring internal representations in transformer models.

Collaborative research identifying and mitigating adversarial attack patterns on ML systems. Part of Trajectory Labs' Apart Martian Hackathon — ranking defensive mechanisms by effectiveness.

Complete installation and configuration guide for Pi-hole on Raspberry Pi and Linux. Network-wide ad blocking, DNS filtering, and privacy hardening for home infrastructure.

Visual, hands-on guide to Docker — containers, image layers, networking, volume mounts, and home-lab deployment. Written for engineers new to containerization.

End-to-end example of containerizing a machine learning application for reproducible, portable deployment — from model training to Dockerized inference service.

Secure, data-governed AI agent on Databricks using Unity Catalog, MLflow, and OpenAI SDK. Implementation from DeepLearning.AI's 'Governing AI Agents' course with extended experiments.

Comprehensive notes and implementations from the AI & Data Science program at Loyalist College. Covers supervised/unsupervised learning, neural nets, evaluation metrics, and production patterns.

A curated collection of advanced Python patterns for machine learning practitioners — generators, decorators, context managers, type hints, and performance tricks.