Bridging Retail Operations, Artificial Intelligence & Independent Research.
I am a Retail Ops Professional transitioning into AI Engineering. I build practical systems that solve real-world problems — from automating stocktake variances to detecting zero-day phishing attacks. I also pursue independent research at the intersection of physics, mathematics, and computation.
Configuration Space, Transcendental Structure, and π-Anomalies at Sub-Planck Scales
Published April 2026 · Zenodo (CERN) · Submitted to Foundations of Physics (Springer)
Key contributions:
- Time is a derived ratio between physical processes, not a fundamental axiom — Newton, Einstein, and Schrödinger recovered from five relational axioms without invoking time
- The three-body problem reduces to geodesic flow on a 2-sphere S² — quantum mechanics solves it exactly where classical mechanics cannot
- The Planck scale is a precision limit, not an ontological wall — π, which defines it, has no final digit
- First base-12 digit search for physical constants in π: electromagnetic constants appear 3.7–4.9× earlier in base-12 π than base-10
- Five original formulas including the Connectivity Force Law
F = −∇(ln κ)and the Geometric Relational HamiltonianĤ_Ω - Mohist relational physics (~400 BCE) and Daoist cosmogony formally mapped to the axiom system
"The universe is a circle. We have been writing its circumference in decimals."
📎 Full paper (open access, 31 pages): https://doi.org/10.5281/zenodo.19371442
After Babel: A Computational Framework for a Universal Written Language Across Human and Artificial Intelligence
Chinese Logographic Tradition, NLP Tokenization, and Blockchain Governance Converge to Solve the Fragmentation of Human Communication
Published April 2026 · Zenodo (CERN) · Preprint
Key contributions:
- The Orthophonemic Coupling Index (OCI = I(G;P) / H(P)) — a formal measure placing any writing system on the spectrum from fully phonographic to fully semantic
- The Channel Separation Theorem (Theorem 7.1): in any writing system where I(G;P) = 0, phonological divergence ΔP has zero path to script divergence — proven via Shannon's mutual information framework
- The Fragmentation Rate Formula
E[ΔG(t)] ∝ OCI₀ · σ² · t— reduces to zero for all time at OCI = 0, making the system mathematically stable against Babel-style fragmentation - The Korean-Japanese natural experiment: Japan retained Kanji (OCI ≈ 0.25) and its script stayed unified; Korea adopted phonographic Hangul (OCI ≈ 0.72) and its written standard diverged within 80 years of political separation
- Egyptian hieroglyphics as the historical warning: started at OCI ≈ 0, corrupted by the rebus principle, became unreadable within a generation — required the Rosetta Stone to decode
- Universal-TF-IDF methodology: combines LaBSE (109-language embeddings) with inverted TF-IDF to mine 400–600 cross-linguistically validated semantic root concepts from multilingual corpora
- A three-level glyph architecture (原符 / 意符 / 合符) structurally isomorphic to modern Byte Pair Encoding tokenization
- Governance framework: 词元 (NLP tokenization) · 代币 (blockchain open standard) · 象征 (semiotics) — extensible without fragmentation, owned by no single culture or institution
"God divided speech at Babel and left song intact. The Universal Script proposes to do for writing what music has done for forty thousand years."
📎 Full paper (open access, 30 pages): https://doi.org/10.5281/zenodo.19416837
- Languages: Python (Pandas, PyTorch, NumPy, Scikit-Learn)
- Engineering: Docker, GitHub Actions (CI/CD), FastAPI, Pytest
- Focus: MLOps, Process Automation, Deep Learning, Cybersecurity
Building robust, modular systems for real-world defense and search.
| Project | Description | Tech Stack |
|---|---|---|
| Phishing Detection System | 🚀 Production Security Engine. A hybrid 3-layer defense system (Whitelist → Heuristic Rules → XGBoost) achieving 99.32% F1-score. Features Real-time Link Expansion, Typosquatting Detection, and a robust CI/CD Pipeline. | Python, XGBoost, GitHub Actions, Pytest |
| Silver Retriever | Offline RAG System. A modular search engine designed for legacy hardware (No GPUs). Features a Plugin Architecture ("The Brain") to detect user intent (Deadlines, Tasks) using TF-IDF instead of heavy LLMs. Includes Smart Chunking. | Python, Streamlit, Scikit-Learn, GitHub Actions |
Engineering intelligent systems that run efficiently on constrained hardware.
| Project | Description | Tech Stack |
|---|---|---|
| Memory Bear (Legacy Edge) | 🐻 Cognitive Agent. A local AI agent running on a 2017 MacBook Air (Intel). Implements Ebbinghaus Forgetting Curves to dynamically manage context window limits. Features Quantized Inference and a biologically inspired Memory Graph. | Python, Llama.cpp, ChromaDB, NetworkX, Phi-3 |
My core focus: Bringing engineering rigor to supermarket logistics.
| Project | Description | Tech Stack |
|---|---|---|
| FreshGuard V2 (Retail Waste) | 🚀 Flagship. Production-grade forecasting engine reducing perishable waste. Features Docker, CI/CD, and Streamlit. The engineered evolution of V1. | Python, Docker, Pytest, Holt-Winters |
| Stocktake Variance Reporter | Automation Tool. A full-stack utility designed to cut stocktake reporting time by 99%. Includes "Theft Detection" logic and a web UI. | FastAPI, Docker, Pandas |
| Enterprise Retail Solution | Advanced R&D. A predictive analytics experiment utilizing Armstrong Cycle Transformers to forecast complex sales demand patterns. | Time-Series, PyTorch Transformers |
| Retail Waste System (V1) | Prototype. My initial menu-driven application for inventory tracking. Focuses on core CRUD operations and basic analytics. | Python, Matplotlib, Pandas |
Applying AI to decode complex genomic sequences.
| Project | Description | Tech Stack |
|---|---|---|
| Genomic Decoder V2 | Advanced Pipeline. Refined Deep Learning architecture for DNA sequencing. Focuses on modular code structure and improved inference performance over the original. | PyTorch, BioPython, CI/CD |
| Genomic Decoder (FlyOS) | Research Implementation. An end-to-end pipeline that reads raw DNA sequences to predict gene expression. Features O(1) lazy-loading for massive datasets. | PyTorch, Transformers |
| Project | Description | Tech Stack |
|---|---|---|
| O-Level Predictor | First App. My very first attempt to convert a Python calculation script into an interactive web app using Streamlit. | Python, Streamlit |