
Pina Studio
A DAW where Claude Code and Codex work in your song with you: they write, arrange, mix and master, and every step lands as an edit you can hear and undo. Written in Rust, drawn on the GPU.
I’m Manyeya, an AI engineer. I build agents and automation that do real work — and interfaces that make them feel effortlessly human.

Agents, music software, and developer tools
built to survive real use.

A DAW where Claude Code and Codex work in your song with you: they write, arrange, mix and master, and every step lands as an edit you can hear and undo. Written in Rust, drawn on the GPU.

An agent-aware terminal multiplexer: coding agents in real terminal panes, and it tells you the moment one needs you.

A download manager that pulls every file in parallel parts, carries on after anything, and takes over your browser's downloads.

Open-source AI coding orchestrator that plans, delegates to sub-agents it can build on the fly, and works in your terminal.

An open canvas for automation — connect your tools, put AI to work, and turn an idea into a workflow that actually runs.

Workflow orchestration engine built on BullMQ.

Local-first prompt engineering environment.
Most AI products fail in the same place — not the model, but the seam between the model and the work. I build the part that closes it: agents and automation that hold up once real people depend on them.
Design and full-stack engineering came first, and that turned out to be the useful preparation. The hard part was never the inference call; it was everything around it. Based in South Africa. Building for people everywhere.
Autonomous agents, LLM workflows and retrieval systems that hold up in production. Tool design, evals and the unglamorous plumbing that decides whether a model is useful or a demo.
Work that runs itself. Connecting the tools a team already uses into orchestrated pipelines, with humans kept in the loop exactly where judgement is worth more than throughput.
The background underneath all of it: years of full-stack architecture and interface design. It is the reason the AI work ships as a product people can use, not a notebook.
Good work is a conversation.
Here’s how I move it forward.
Find the work worth automating. Which steps are repetitive, which need judgement, and where a model earns its keep instead of adding a new failure mode.
A narrow end-to-end slice, fast. Real data, real prompts, real interface — enough to see whether the idea survives contact with how people actually work.
Type-safe systems around the model: tool boundaries, retries, guardrails and an interface that makes the AI legible rather than mysterious.
Evals, tracing and cost monitoring after launch. Model behaviour drifts; the system that watches for it is part of the product, not an afterthought.
Interface studies from the design years. Still the reason the AI work ships looking like a product.
Thoughtful projects start with a conversation.