Software intelligence / research programme
Codepot
Make software structure explicit enough for humans and AI to reason about it together.
Codepot explores typed DSLs, compiler pipelines, contract-first development and constrained generation. The premise is simple: AI becomes more reliable when it works inside a system that can validate what “correct” means.
galaxy Commerce
construct Order {
id: uuid
total: currency<USD>
status: OrderStatus
}
representation OrderAPI {
expose Order
validate before write
}
# structure becomes
# executable contextResearch thesis
Structure before generation.
Language
A compact typed language that captures intent without collapsing into implementation detail.
Compiler
A deterministic pipeline that can parse, validate and lower intent into reusable intermediate representations.
Tooling
LSP, diagnostics, autocomplete and navigation that make the language practical for real engineering work.
Evaluation
Generated outputs are checked against explicit contracts rather than judged only by whether they look plausible.
From contract to system
The compiler is not just a code generator. It is the boundary between intent and implementation.
01
Intent
Human-authored domain and system definitions.
02
Parse
Syntax becomes a typed syntax tree.
03
Validate
Names, types and contracts are resolved.
04
Lower
A stable IR becomes the source for tooling.
05
Generate
Targets are produced inside known constraints.
Current questions
What we are trying to learn.
How much software intent can be expressed before a system becomes too verbose to use?
Can generated code remain replaceable and reviewable instead of becoming opaque AI output?
What should an AI agent be allowed to change when a contract is the authority?
Can the same typed source drive code generation, documentation, validation and developer tooling?