Independent AI researcher with a degree in Philosophy and Theology (understanding the presuppositions of science),currently completing coursework for a double major in Physics and Computer Science.
Areas of Interests: metaphysics, physics of intelligence, quantum information science, activation steering, behavioral monitoring, philosophy of mind and language, and the mathematical foundations of neural computation.
scale-invariant-interpretability Investigating whether mathematical structures in small neural networks predict behaviors in larger models. Proposes that mathematical invariants computed from small models remain conserved at scale, enabling efficient cross-scale analysis.
persona-vector-probes Examining whether internal model vectors correlate with reported phenomenological experience. Tests the relationship between mechanical behavior modification and subjective states through behavioral proxies.In other words, I am testing if activation vectors correlate with subjective experience. Thus, answering: Can we measure what models 'feel'?" Framework designed to generate useful data regardless of outcome.
avat-behavioral-steering Testing whether activation steering can induce instrumental misalignment behaviors through vector manipulation. Builds on activation addition research (Turner et al., 2023) and contrastive activation steering (Rimsky et al., 2023). Safety research identifying vulnerabilities to develop defenses.
research-notes Critical analysis of fundamental limitations in current AI alignment approaches. Identifies three core constraints and fourteen failure modes across Constitutional AI, RLHF, and interpretability techniques. Engages with counter-evidence and frames claims as hypotheses worth investigating.
principiadynamica | PyPI: constitutional-dynamics State Transition Calculus (STC) for modeling AI alignment as continuous trajectory through embedding space. Monitors behavioral drift via φ-alignment scores, Lyapunov stability, and frequency-domain analysis. Extends Constitutional AI to runtime monitoring.
kernel-ridge-steering Examining whether spectral properties of activations predict effective steering intervention points. Early empirical results on GPT2 showing correlation (r=0.61) between spectral gap and steering effectiveness.
ctxpack LLM context compression. Aims to achieve reduction through orthogonal storage and query compressions. Employs a Degree system (1-10) that maps to cache tiers from seed summaries to full sessions. Compress once, query at any resolution.
graph-aware-rag Retrieval system using Code Property Graphs with multi-modal attention and information-theoretic ranking. Aims to achieve better context selection than traditional RAG by understanding code structure.
codeviz Code analysis bridging static understanding with conversational context for LLM-optimized prompts. Multi-language AST parsing, semantic embeddings, token-aware generation.
personalized-news-bot Multi-agent system transforming news into personalized intelligence through 7-agent pipeline. Evolved through 6 prompt engineering iterations with rigorous evaluation. Key finding: when prompt engineering succeeds vs when architectural solutions are essential.
llm-playground Research monorepo with 10+ specialized LLM agents, multi-agent orchestration, constitutional-debate framework, and MCP server hosting.
XFN-CFPE (Prompt Engineering Infrastructure) Closed-loop system for systematic prompt engineering at scale. Hypothesis-driven development with execution engine, evaluation rubrics, and data-driven iteration. Contains DIALECTICA system prompts, multi-provider response collection, and comparison harnesses for prompt optimization and safety testing.
mechanistic-interpretability-techniques Tracks a significant amount of papers covering 23 mechanistic interpretability techniques. Summarizes paper main finding alongside its working implementations. Learning through building. Favors breadth over depth.
transmuter-project Aim: Python interpreter built from scratch. Tackles metaclass magic, descriptor protocol, MRO, generators, and other features that break toy interpreters. Classic architecture: Lexer → Parser → Evaluator → Object System.
arm_sorter Raspberry Pi robotic arm with RFID and vision integration. Built physical arm, wiring scaffolds, servo control. Designed MCP server for Claude integration. Currently paused.
puzzles Logic puzzles when time permits.To date features 1 Jane Street challenge, philosophical logic in Z3.