Verified improvements on functions from open source numerical software projects. Correctness checked end-to-end.
Connect your GitHub repository. DeltaCode clones your code into an isolated container, scans it for reachable functions, benchmarks their runtime with real inputs, and returns pull requests with faster equivalents.
DeltaCode scans your repository to identify the functions reachable from your entry scripts. You choose which ones to target for optimisation.
Baseline runtime is measured on real inputs. GPT-4o generates faster equivalents, which are executed and benchmarked in parallel containers.
Each accepted optimisation ships as a GitHub pull request. The diff, benchmarks, test outputs and full audit trail are included for review.
Your code runs in disposable containers, spun up per job and destroyed on completion. No shared state, no persistent execution environment, no lateral access between clients.
Each optimisation job spins up a fresh Azure Container Instance for the duration of the run. When the job completes, the container is destroyed. Nothing persists between runs.
Every client has isolated compute, storage and access tokens. No client's data or code can ever be seen by another. Data isolation is enforced end to end at every layer of the stack.
GitHub access tokens are stored encrypted with Fernet keys. They are only decrypted at the moment of an optimisation job and never logged or exposed to third-party services.
Only optimisation results, benchmarks and metadata are retained for reporting and audit. Full source code is not stored beyond the duration of an active job.
The LLM generates candidates. A deterministic verification layer decides which ones ship. Every optimisation is executed end-to-end against real test inputs before it is ever proposed as a pull request.
The LLM proposes candidate implementations within a defined scope. It never controls execution, never touches the deployment pipeline and never runs unbounded.
Every candidate is executed against real inputs collected from your entry scripts. Outputs must match the baseline byte-for-byte before an optimisation is considered valid.
pull requests are only opened when the measured improvement passes your configured threshold. Marginal or noisy speedups are rejected automatically.
DeltaCode fits into the workflow your engineering team already uses. Optimisations arrive as GitHub pull requests, ready for your normal code review process.
Connect any GitHub repository. DeltaCode uses standard OAuth authentication and opens pull requests that appear in your usual review queue.
Optimised code passes through your existing CI/CD pipeline like any other commit. Your existing checks, review policies and merge rules apply.
Configure optimisations to run automatically on push, on pull request, or on a schedule. Or trigger them manually when you want to.
Every codebase is different. Every engagement is quoted individually, based on the shape of the problem and the scope of the work.
Every engagement is scoped in an initial call. Pricing depends on codebase size, complexity of dependencies, entry point configuration, and the depth of optimisation coverage you need.