Joe Munene
Email: joemunene984@gmail.comContact me
Available for Work
Hi! I'mJoe Muneneasecurity and AI engineerbasedin Nairobibuilding secure systems, from the model up.

Founder of Complex Developers. I build security-focused AI from scratch: GhostLM, an 81M-parameter language model trained in pure PyTorch, and ghostloop, a fail-closed agent runtime on PyPI. Across 50+ repositories, with upstream contributions to PyTorch Ignite, AutoGPT, and OWASP.

See what I can do
Shipped on PyPI*Live on the App Store and Google Play*Models on Hugging Face*Merged into AutoGPT*SARIF in CI*Shipped on PyPI*Live on the App Store and Google Play*Models on Hugging Face*Merged into AutoGPT*SARIF in CI*
{01} / Featured projects

Real tools, shipped and in use.

Get in touch
AI / ML81M params, pure PyTorch

GhostLM

An open-source cybersecurity language model built from scratch in PyTorch (RoPE, SwiGLU, RMSNorm), trained on a 422M-token corpus, with a tool-using agent runtime, multi-vendor API and MCP servers, and a rigorous eval suite.

GhostLM
Source
AI / MLPyPI v1.0.3, 359 tests

ghostloop

An agent runtime and fail-closed safety pipeline for embodied AI, published on PyPI with a live demo, a Next.js dashboard, and a Tauri desktop app. Six backends, twelve policy gates, and a statistically rigorous bench harness.

ghostloop
Source Live demo
macOS / MLSwift, Metal, Core ML

ghostscale

Real-time neural upscaling for Windows games on Apple Silicon. ScreenCaptureKit capture, a per-game GhostNet model on the Neural Engine with Metal-side pixel shuffles, and a fullscreen overlay: 33.7 dB mean PSNR against native 1080p versus MetalFX's 26.4 to 33.5, at 54 fps in live play.

ghostscale screenshot
SecurityMCP, policy gates, sandbox

secure-mcp

An MCP server that exposes security tools to AI agents through a fail-closed policy pipeline: subprocess sandboxing, per-call audit trails, and YAML authorisation gates. Drop-in for any agent that runs offensive tooling.

policies/default.yaml
tools:
  port_scan:
    enabled: true
    target_allowlist:
      - "127.0.0.1"
      - "10.0.0.0/8"
    max_ports: 100
    rate_limit:
      per_minute: 10
Source
Web / SaaSNext.js, Stripe, PostHog

ChartSentinel

Trading-intelligence SaaS: chart-event alerting and signal aggregation for retail and prop traders. Next.js and TypeScript with a Stripe paywall, PostHog analytics, Sentry, and Resend, shipped end to end.

ChartSentinel screenshot
Live Source
Mobile & webiOS, Android and web Verified

EzCare

Lead mobile and web developer on EzCare, an AI wellness companion live on the App Store, Google Play and the web: React Native and Expo shipped through EAS, a Supabase backend with a tRPC data layer, RN Skia and Reanimated, and the EZBuddy AI coach on the Anthropic SDK, all in one Bun and Turborepo monorepo.

EzCare screenshot
App Store Google Play Source
Desktopv0.7.1, Tauri + Rust

linkdrop

An iPhone-to-Linux companion app: browse photos and files, see notifications, and mirror the screen. Cross-platform Tauri and Rust build with a daemon-backed bridge and CI-built .deb and .AppImage releases.

linkdrop screenshot
Source
View all projects
{02} / Tools & Skills

My technical toolbox

Python
Language
Expert
TypeScript
Language
Expert
Rust
Language
Proficient
JavaScript
Language
Expert
Go
Language
Proficient
C / C++
Systems
Proficient
Swift
macOS / Apple
Proficient
Shell / Bash
Scripting
Advanced
HTML / CSS
Web
Advanced
React / Next.js
Web
Expert
Node / FastAPI
Backend
Advanced
PyTorch
AI / ML
Advanced
From-scratch transformers
AI / ML
Advanced
React Native / Expo
Mobile
Advanced
Tauri (Rust)
Desktop
Proficient
Penetration testing
Security
Advanced
MCP / agent safety
AI security
Expert
Docker / Kubernetes
DevOps
Advanced
Git / GitHub
Tooling
Expert
Linux
Systems
Advanced
Red team, blue team*Models, runtimes*Web, mobile*Desktop, robotics*Red team, blue team*Models, runtimes*Web, mobile*Desktop, robotics*
{03} / Experience

Working Experience

Founder/01

Founder & Lead Developer, Complex Developers, Nairobi

  • Studio: Run it end to end across client work, architecture, engineering, and deployment.
  • Platform: Built the company platform on Next.js and Prisma with a Postgres admin CRM that syncs flagged projects to the public site.
  • Standards: Standardised the stack (Vercel, Supabase, JWT auth, shared design tokens) across every project.
2026 - Present
Contract/02

Lead Mobile & Web Developer, EzCare AI (contract, remote)

  • Platforms: Lead development of EzCare on iOS, Android and the web from one Bun and Turborepo monorepo, shipping releases through EAS.
  • Product: Supabase backend with a tRPC data layer, the EZBuddy AI coach on the Anthropic SDK, onboarding, Google and Apple sign-in, and the paywall.
  • Verified: Engagement confirmed by a certificate signed by the founder of EzCare AI.
Jan 2026 - Present
Independent/03

Independent Developer & Security Researcher, Self-Employed, Nairobi

  • 50+ projects: Shipped across security, AI and machine learning, and full stack.
  • GhostLM: An 81M-parameter transformer built from scratch in PyTorch on a 422M-token corpus, with GhostBench and an MCP server.
  • ghostloop: A fail-closed agent runtime on PyPI (v1.0.3) with 359 tests, a live demo, a web dashboard, and a desktop app.
  • Desktop: linkdrop, a Tauri and Rust app bridging an iPhone to Linux.
2023 - Present
Open source/04

Open source: AutoGPT, PyTorch, PyTorch Ignite, OWASP

  • AutoGPT: Fixed agent-name preservation in the visual builder (#12805, merged).
  • PyTorch: Per-FQN CPU offload and budget-based offload for FSDP2 (#188120, draft PR in review).
  • PyTorch Ignite: Welford-algorithm numerical-stability fix for PearsonCorrelation (#3741, open PR).
  • OWASP: Proposed an AI Model Supply Chain Security cheat sheet, revised through maintainer review.
2026
81M
Parameter LM
trained from scratch in PyTorch
1
Merged upstream
AutoGPT, plus open PRs to PyTorch
Live
Shipped to production
PyPI package + iOS and Android apps
{04} / How I work

Building real systems from the model up.

* A polymath by intent: security, machine learning, robotics and full-stack, all built from scratch, tested and shipped.

{05} / Currently building

What is on the bench right now

GhostLM v0.9.33 → ghost-base

Ghost-small (81M) has hit its parameter ceiling, confirmed empirically across the GhostBench suite. The next move is the rented-GPU ghost-base run at roughly 360M parameters on the new 422M-token corpus.

92% complete
  • ✓From-scratch decoder transformer (RoPE + SwiGLU + RMSNorm)
  • ✓v1.0 corpus build (422M tokens, 27 sources, leakage-clean)
  • ✓12 differentiation bets + GhostBench eval suite
  • ✓GhostAgent runtime + multi-vendor HTTP / MCP server
  • ✓Open-source code corpus pull (105 repos, 168M chars)
  • 6Ghost-base v1.0 GPU run (rented H100, ~360M params)
  • 7Acceptance gate: ≥40% CTIBench OR ≥65% CTF eval
{06} / The ghost suite

Eleven security tools, one suite.

View the suite

Offense, defense, supply chain, cloud and crypto. 489 passing tests, CI green on Python 3.11 and 3.12, and SARIF output so every tool drops into code scanning.

ghostreconOSINTPassive recon: cert transparency, DNS, Wayback, entity graph and reports.ghostmapAppSecWeb vulnerability scanner with an auth-aware crawler and XSS and SQLi engine.ghostpwnOrchestrationPentest workflow engine: YAML stages run as a parallel dependency DAG.ghostscopeDetectionML intrusion detection with IsolationForest and a PyTorch autoencoder.ghostboxMalwareStatic malware triage: PE and ELF parsing, IOCs, YARA and a threat score.ghostdlpDataLeak prevention with 16 validated PII and secret detectors and redaction.ghostsbomSupply chainCycloneDX SBOMs, OSV.dev CVE scans and typosquat detection.ghostchainWeb3Solidity auditor with 10 SWC-mapped detectors and SARIF output.ghostmobileMobileAPK and IPA analyzer with a dependency-free binary manifest decoder.ghostcloudCloud17 AWS, GCP and Azure misconfiguration checks with remediation.ghostvaultCryptoKey management with envelope encryption, rotation and an audit log.
{07} / Qualifications

Education

01/B.Sc. Computer Science, Moi University+
2025 - Present, Nairobi. Coursework in data structures and algorithms, operating systems, computer networks, and discrete mathematics. Self-directed focus on offensive security, transformer architectures, and systems programming.
02/High School Diploma, Nkaimurunya Mixed High School+
Ongata Rongai, Kenya. Secondary education, where the interest in computing and self-directed technical learning started.
03/Self-directed study, security, ML & systems+
From-scratch language model engineering (attention, tokenization, training), CVE reasoning and exploit-chain analysis, active CTF competition, and open-source contributions to PyTorch Ignite, AutoGPT, and OWASP.
{08} / Contact me

Feel free to connect, explore my work, or reach out directly.

GitHubHugging FaceDev.to
Phone+254 718 733 968+254 116 549 448
EmailResume / CVGet in touch
Open to freelance*Open to full-time roles*Based in Nairobi*Working remotely*Open to freelance*Open to full-time roles*Based in Nairobi*Working remotely*
Available for workBack to top

Let's connect and explore future opportunities.

Let's make an impact.

Joe Munene

Joe Munene

Security and AI engineer, Nairobi

Contact me
joemunene984@gmail.com+254 718 733 968+254 116 549 448

Feel free to reach out if you would like to work together on security, AI, or software.

Reach out
Joe Munene
Copyright 2026 Joe Munene. All rights reserved.