Product strategy / 01
A semantic layer.
An infrastructure business.
Semryn should help coding agents understand and change existing software with less repeated discovery, more explicit dependencies and stronger validation.
Direction: Semryn Context Engine, powered by AIR · Status: early prototype
The product thesis
Make AI coding agents significantly cheaper, faster and safer on real software repositories.
This is the outcome we intend to prove in pilots. Semryn is building infrastructure between coding agents and repositories. Customers should continue using TypeScript, JavaScript, PHP, Python, Java, Rust, C# and their existing tools. No application rewrite is required.
Semryn is designed to complement Codex, Claude Code, Copilot, Gemini and custom agents. Those integrations are a product direction; they are not delivered by the current prototype.
Product and technology
Semryn is the product/platform. AIR — AI Intermediate Representation — is the semantic representation and compiler technology underneath it. AIR graphs describe types, dependencies, effects and constraints; Semryn is the infrastructure being built around them.
The AIR specification and compiler are candidates for an open-source foundation. Licensing and distribution are not finalized; no license is asserted here.
From repeated discovery to reusable understanding
- Index the repository. Extract symbols, dependencies and source references at a known revision.
- Select useful context. Return a bounded view of the code and tests relevant to a particular task.
- Propose a semantic patch. Address an intended entity with explicit preconditions.
- Validate the change. Check graph rules, source revisions and the repository’s language, build and task checks.
- Produce a minimal source diff. Preserve unaffected source and the team’s normal review process.
Prove the economics
Compare the same tasks with and without Semryn using the same agent, model, repository revision and acceptance criteria. Count cold indexing, warm queries, retries, repairs, validation and unsuccessful attempts.
| Outcome | What we will measure |
|---|---|
| Less context | Model-visible tokens, relevant coverage and context quality |
| Less repeated work | Total tool calls, failed attempts and repair cycles per accepted task |
| Safer changes | Escaped regressions, failed checks, stale edits and unintended scope |
| Lower total cost | Inference, indexing, engine and validation costs per accepted task |
| Faster delivery | End-to-end task time, including setup and verification |
Compact binary transport and fast graph operations are useful engineering properties. They do not establish a universal reduction in model tokens, task time or inference cost.
The next milestones
01 — Semryn repository intelligence
One existing source language, one agent adapter and a curated set of repository tasks. Demonstrate useful, fresh context and measure cold and warm overhead against the existing workflow.
02 — Semryn agent infrastructure
A narrow edit set, explicit source preconditions, source-preserving lowering and parser/build/test checks. Accepted output remains a normal Git diff.
03 — Commercial evidence
Repeatable benchmarks and real team trials. Establish customer-defined quality and cost thresholds before expanding support or making performance promises.
04 — A dependable product
Index lifecycle, policies, audit receipts and deployment support based on actual customer requirements. Expand language and agent coverage where demand and measured economics justify it.
Semryn’s current AIR foundation
Working in AIR: binary native/web semantic graphs, typed construction, validation, checked attribute edits, persistent local binary control, demonstrated native and static HTML outputs, and host-owned checks.
Planned: existing-repository indexing, task-specific context extraction, source-language adapters, semantic source patches and external coding-agent integrations.
Validation can establish the properties and observable requirements it checks. It does not guarantee correct business logic or eliminate every AI mistake. Human review and meaningful task checks remain part of the intended workflow.
Download the full strategy