Vision before execution.
Design before code.
The story of josomu, why the "vibe coding" label misses the mark, and what it actually means to operate as an AI-Native Product Engineer.
Why syntax is cheap, but taste is everything.
When generative AI first exploded, people coined "vibe coding" to describe throwing casual prompts at an LLM and hoping for something usable. While entertaining, building real, delightful, production-grade products requires something far more rigorous.
Code syntax has become commoditized. You can generate a thousand lines of boilerplate in seconds. But syntax was never the heart of engineering — understanding user mental models, architecting clean state machines, optimizing latency budgets, and obsessing over interaction feel are.
As an AI-Native Product Engineer, I don't surrender intent to an algorithm. I use autonomous multi-agent systems and fine-tuned local models as high-speed execution engines to bring deliberate, human-designed concepts into existence with unprecedented momentum.
// The vibe is NOT the spec.
// The spec is the vision, the system architecture, and the emotional resonance.
How we build at josomu
Our engineering pipeline translates raw curiosity into polished systems through a structured 4-step loop:
Mapping out user psychology, edge cases, data flows, and spatial wireframes before a single agent touches the codebase.
Deploying autonomous coding harnesses (Antigravity, Codex, Claude Code) along with private local reasoning engines (Qwen, Ornith) across shared context bridges to implement features and run sandboxed tests concurrently.
Rigorous benchmarking on local hardware. Strict privacy controls, zero dependency bloat, and fast, offline-first reliability.
The critical final mile: tuning physics curves, typography pacing, micro-animations, tactile sound feedback, and UI delight that no machine can feel.
Technologies & Core Stack
We choose technologies optimized for speed, sensory delight, and system reliability:
AI & Agent Orchestration
Google Antigravity, OpenAI Codex, Claude Code, custom context memory bridges, and privacy-preserving local models (Qwen, Ornith).
3D, Vision & Neural Pipelines
TRELLIS.2 neural 3D asset generation, Three.js, WebGL, MediaPipe (on-device vision), Unity / C#, procedural geometry.
Web & Spatial Systems
TypeScript, React, Vite, Supabase / Postgres (Realtime & RLS), Node.js, Canvas APIs.
Native & DevTools
macOS system utilities, Python local pipelines, Android (Kotlin), Google Play Console.
Support the Studio
josomu is entirely self-funded and independent. If you enjoy our playable games, open-source utilities, or design research, you can support future developments directly on Ko-fi.