Joe Munene

• Lead development of EzCare across iOS, Android and the web, an AI wellness companion live on the App Store and Google Play.
• Built the mobile app in React Native and Expo and ship its releases through EAS Build and Submit; the web app shares code with it in a Bun and Turborepo monorepo.
• Designed the Supabase backend (PostgreSQL, auth, storage) with a type-safe tRPC and TanStack Query data layer, and built the EZBuddy AI coach on the Anthropic SDK.
• Implemented onboarding, Google and Apple sign-in, the subscription paywall, and native camera, haptics and voice features.

• Run the studio end to end, covering client work, architecture, engineering, and deployment.
• Built the company platform on Next.js and Prisma with a Postgres-backed admin CRM, where flagged projects sync to the public site automatically through Server Components and incremental static regeneration.
• Standardized the stack across projects: Vercel and Supabase for deploys, JWT-authenticated API routes, and a shared design token system that keeps the brand consistent from landing page to dashboard.

• Shipped 50+ open source projects across security, AI and machine learning, and full stack.
• Built GhostLM, an 81M parameter decoder-only transformer, entirely from scratch in PyTorch with RoPE, SwiGLU, and RMSNorm, without using the transformers library. Trained it on a 422M token corpus I collected across 27 cybersecurity, code, and reasoning sources.
• Shipped GhostBench alongside it, an evaluation suite using proper statistics like Wilson confidence intervals and McNemar paired tests, plus GhostAgent, a tool-using runtime, and an MCP server that works with Claude Desktop and Cursor.
• Merged a contribution upstream into AutoGPT (#12805), fixing agent-name preservation in the visual builder's AgentExecutor node titles after reload.
• Proposed a numerical-stability fix to PyTorch Ignite (#3741, open PR): replacing an unstable variance formula in PearsonCorrelation with Welford's algorithm, which returned wrong results on large-mean float32 data.
• Built ghostloop, an agent runtime and fail-closed safety pipeline for robotics, published on PyPI (v1.0.3) with a live demo and full CI/CD, 359 passing tests, and extended into a web dashboard and a Tauri desktop app.
• Shipped linkdrop, a Tauri and Rust desktop app that bridges an iPhone to Linux for photos, files, and screen mirroring, with CI-built .deb and .AppImage releases.
• Built a cybersecurity tooling set including an MCP security server with sandboxing, a Kubernetes CIS auditor, a memory forensics tool, a SIGMA-rule SIEM, and a pre-commit secret scanner.
• Wrote the AI Model Supply Chain Security guide, proposed to the OWASP Cheat Sheet Series and revised through maintainer review.

Coursework in data structures and algorithms, operating systems, computer networks, and discrete mathematics. Self-directed focus on offensive security, transformer architectures, and systems programming.