Ashwin Govind Ugale is a software engineer at Enidus USA who builds AI-backed telecom workflows and evaluation tools. His muteval project tests whether an evaluation suite notices deliberately introduced regressions.[3]
Start with Muteval – mutation testing for your LLM evals →
Education
B.Tech in Electronics and Telecommunication
Aug 2019 – May 2023
[3]Links
Software Engineer · Enidus USA
Career
Ugale’s public résumé describes a multi-tenant tool-calling platform at Enidus for telecom reseller operations, including tenant isolation, persistent workflow state and approval before state-changing carrier actions. Earlier roles include software engineering at Live Risk Labs and an internship at Hyphenova.[3]
Contributions
Test the tests by introducing failures
muteval deliberately changes a system under test, reruns its existing evaluations and reports which injected regressions were caught. A surviving mutant indicates a possible coverage gap; an unchanged output can instead mean that the mutation was inert. The project distinguishes these cases and offers candidate evaluations for review. Its mutation score measures sensitivity to the selected perturbations, rather than establishing overall application correctness or measuring every possible regression.[2]
Ground answers in the dependency version actually installed
DocChat reads a project’s lockfiles and routes retrieval to documentation indexed for the pinned library versions. It returns citations with an answer and can refuse questions outside the indexed material. Its README also documents a circular refusal metric that was corrected by using independently labeled scope. That correction shows why version grounding and evaluation design need separate checks; a plausible answer or a high headline score is insufficient.[5]
Tools, writing & talks
Writing, research & talks
Start here
Mutation testing for LLM evaluation suites, with survivor reports and inert-mutation handling.
Run the documented offline demo, then inspect a surviving mutation and decide whether it represents a failure your application should catch. Python and an existing evaluation interface are needed for integration; suggested new tests need human review.
Repository owner and developer[2]Original source