joe munene
@Joe Munene
Joe Munene
AI researcher & full-stack engineer | Building LLMs from scratch
About
Software engineer focused on secure systems, full-stack development, and applied cybersecurity.
Top Skills
Python
•
TypeScript
•
JavaScript
•
Rust
•
C
•
SQL
•
PyTorch
•
Machine Learning
•
LLM Research
•
Natural Language Processing
•
Deep Learning
•
Penetration Testing
•
Cybersecurity
•
OSINT
•
Threat Intelligence
•
React
•
Next.js
•
Node.js
•
FastAPI
•
REST APIs
•
Docker
•
Linux
•
Git
•
CI/CD
•
Cloud Infrastructure.
Experience
work
Software Engineering Intern
Besechi
Mar 2026 - Apr 2026Gained hands-on industry experience in software development. Contributed to real production systems in a professional engineering environment
work
Founder & Lead Engineer
Complex Developers
Jun 2025 - PresentBuilt and led a software agency offering full-stack development, AI integration, cybersecurity consulting, and cloud infrastructure. Delivered client projects spanning CRM platforms and BPO systems.
Education
school
MOI University
Bachelor, Computer Science
2025 - 2028Computer Science program covering software engineering, database systems, networking, and applied development. Concurrent with active open-source research and agency work.
Projects
GhostLM
Visit Project open_in_new Apr 2026 - PresentOpen-source language model built from scratch in PyTorch. Purpose-built for cybersecurity, with code + general language + math reasoning folded into the v1.0 corpus (516K records, ~363M tokens, six domains).
Language & Tools
Languages
TypeScript, JavaScript, Python, Rust, C, SQL, Shell/Bash, HCL, HTML/CSS
Frameworks
Next.js 15, React 19, Node.js, Tauri, PyTorch, Prisma, PostgreSQL, FastAPI, Express, Jupyter, Vite
Security
Penetration testing, vulnerability assessment, CTF, network analysis, OSINT, memory forensics (YARA/Volatility), SIGMA/SIEM, Kubernetes CIS auditing, AI-agent sandboxing, steganography
Tools
Git, Linux, Docker, Terraform, Vercel, Supabase, Nmap, Wireshark, Burp Suite, Metasploit, YARA, Volatility
AI / ML
Transformer architectures, attention mechanisms, from-scratch LLM training, tokenization (BPE), MCP servers, AI-agent policy / sandboxing, LLM integration, Claude API, Groq
TypeScript, JavaScript, Python, Rust, C, SQL, Shell/Bash, HCL, HTML/CSS
Frameworks
Next.js 15, React 19, Node.js, Tauri, PyTorch, Prisma, PostgreSQL, FastAPI, Express, Jupyter, Vite
Security
Penetration testing, vulnerability assessment, CTF, network analysis, OSINT, memory forensics (YARA/Volatility), SIGMA/SIEM, Kubernetes CIS auditing, AI-agent sandboxing, steganography
Tools
Git, Linux, Docker, Terraform, Vercel, Supabase, Nmap, Wireshark, Burp Suite, Metasploit, YARA, Volatility
AI / ML
Transformer architectures, attention mechanisms, from-scratch LLM training, tokenization (BPE), MCP servers, AI-agent policy / sandboxing, LLM integration, Claude API, Groq
Currently Exploring
GhostLM
An open-source cybersecurity-focused language model built from scratch in PyTorch. Decoder-only transformer trained 10K steps on CVEs, CTF writeups, and security research.
An open-source cybersecurity-focused language model built from scratch in PyTorch. Decoder-only transformer trained 10K steps on CVEs, CTF writeups, and security research.
Achievements
Upstream contributor to PyTorch Ignite and OWASP. Published research on CTIBench evaluation bias. Built and shipped ghostloop to PyPI and HuggingFace.
Fun Fact
I can't learn something without wanting to understand how it works at the lowest level possible.
Random Dev Quote
The best way to understand a system is to build it yourself and watch it fail.
Contributions
Comments
- @[Flybywire] Thanks, appreciate it. The trickiness is exactly why I w...Jun 12, 2026
- @[sumita] Thanks, and that is the hardest question in the whole thing,...Jun 12, 2026
- @[sumita] Great question, and it is the exact failure mode I was worri...Jun 12, 2026
- @[peculiarlibrarian] Exactly, and the challenge is that most threat in...Jun 12, 2026
- This intersection becomes fascinating in cybersecurity, mapping relati...Jun 12, 2026
Articles
- I built a red-team scanner for MCP servers. Then I pointed it at the real ones.Jun 12, 2026
- I thought my LLM hit 36.9% on a cybersecurity benchmark. It was picking C every time.May 11, 2026
- What I Learned Building a Cybersecurity LLM From Scratch: 30,000 Steps, 3 Architecture Rewrites, andMay 02, 2026
- I found a bug that made my LLM look 14x better than it was — here's what I learnedMay 02, 2026
chevron_left