MigrationAssurance

Evidence-grade sign-off for model swaps and agent deployments.

Model deprecations force migrations on someone else's schedule. "We eyeballed ten outputs" is not evidence a regulated enterprise can take to a risk committee. This site is about migration assurance: the evals, statistics, and governance mapping that let you say "the new model is safe to deploy" and prove it.

Work

signoff, an open-source LLM migration assurance harness: N-run, nondeterminism-aware A/B equivalence testing with layered matching (exact, normalized, numeric, then LLM-judge), flip-rate statistics, and an audit-ready report. Vendor-neutral by design.

nightwing, a reproducible benchmark of frontier LLMs vs a cheap fine-tuned specialist on CUAD contract clause extraction, scored on the official metric with every model's raw predictions published. Version two changed one thing, the task framing, and gained ten points: the specialist now beats GPT-4o in 25 of 40 categories and GPT-5.2 in 22 of 40, still loses overall to Claude, and its pre-registered prediction missed in public. Honest numbers over highlight reels. The trained model is on Hugging Face.

Writing

About

Twelve years ago I trained my first ML model. Today I'm Ashish Kumar Singh, an engineering leader building the LLM systems enterprises trust with decisions worth millions: contract intelligence now, a decade of clinical AI in healthcare before that. 100+ models in production at 99% uptime. Fraud detection across 10 million insurance claims. Behaviour-change AI reaching 2 million people a day.

And I still write the code. Migration validation, prompt regression testing, document-extraction evals: the unglamorous discipline that decides whether an AI system survives contact with production. That discipline has a name, migration assurance, and this site is where I write about it. signoff is its open-source toolkit.

Elsewhere: GitHub · LinkedIn.

All views on this site are my own. I write from personal experiments and open-source work, never from client or employer data.