An LLM authors the artifact at compile time. The runtime stays free of inference.
Compiled AI is a paradigm for systems that must be deterministic, auditable, and fast. The model runs offline to draft a config, rule, or policy; an external validator proves the artifact does what it claims; the committed artifact is then consumed by a boring, well-known tool — Semgrep, Conftest/OPA, OpenFisca, OHDSI — with no model in the request path.
Start with compiled-ai — the paradigm and its five-part shape: spec → compiler → gates → artifact → runtime.
Same shape, four domains. The compile input and the deterministic runtime change; the gated-compile discipline does not. Each commits its artifact only after machine-checkable gates pass, and treats refusal — where the input admits no single determinate encoding — as a first-class output, never a guess.
| # | Repo | Compiles | Into | Runtime |
|---|---|---|---|---|
| 1 | semgrep-rule-compiler | incident postmortems + code samples | Semgrep rules | Semgrep |
| 2 | terraform-policy-compiler | plain-English Terraform policies | Conftest/OPA Rego | Conftest / OPA |
| 3 | tax-rules-compiler | statutory income-tax text | an OpenFisca package | OpenFisca |
| 4 | trial-eligibility-compiler | verbatim ClinicalTrials.gov eligibility criteria | an OHDSI Circe cohort | OHDSI (CirceR) |
- Use it. Open an issue with a pattern you'd want compiled.
- Extend it. PRs for new languages, rule modes, or autofix — all gated the same way.
- Build the next compiler. Have a deterministic tool that consumes config or rules (GitHub Actions, Dependabot, an in-house policy engine)? Open an issue describing the use case — we're looking for collaborators on adjacent compile targets.
- Essay: "Compiled AI: Engineering Deterministic LLM Systems" — Boris Teplitsky, ITNEXT, 2026.
- Related paper: Trooskens et al., "Compiled AI: Deterministic Code Generation for LLM-Based Workflow Automation" — arXiv:2604.05150.
Contact: open an issue on any repo, or [email protected].