Project Semilla
Active Research • Research Initiative
A self-evolving software platform that can inspect, modify, test, and improve source code while preserving user ownership and control.
WHY IT EXISTS
Most AI coding tools work from the outside: a model suggests code, a person applies it. Semilla asks a harder question — can software participate in its own development? A requested capability is translated into development tasks, code is generated or modified, the result is compiled, failures are inspected, corrections are attempted, and successful changes are integrated. The long-term question is not simply whether AI can generate code, but whether software can evolve while remaining understandable, controllable, and owned by the person using it.
HOW IT WORKS
The evolution loop: user requirement → development tasks → generate or modify code → compile → inspect feedback → correct → integrate. The system is built in C# with Roslyn for dynamic compilation, using local LLMs (Ollama, DeepSeek) so the intelligence layer runs under the user's control.
WORKING NOW
The functional proof of concept generates working modules — including user-interface components — with compilation and feedback inside the loop, and runs on macOS, Windows, and Linux.
RESEARCH DIRECTION
Path handling, broader testing, refinement of the main GUI, review of GitHub integration, and stronger user control over the evolution process. The goal is persistent capability accumulation without giving up ownership or oversight.