Findings from
our work.
Technical articles, evaluation reports, and model notes from Navitra — what we trained, how we measured it, and what we learnt. Written for practitioners, not press releases.
Evaluation reports
Benchmark results and held-out test metrics for models we have trained — with methodology, dataset composition, and known limitations included.
Technical notes
Architecture decisions, training observations, and deployment patterns from real engagements — the things that do not make it into a product page.
Domain reviews
Analysis of how AI is being applied in specific sectors — finance, life sciences, infrastructure — and where the genuine open problems remain.
Models we have built.
A code generation model trained to produce correct, idiomatic output across multiple programming languages. Optimised for developer workflows and automated testing pipelines.
View model page →A 149.6M-parameter decoder-only protein language model trained from scratch on 2.45M UniProtKB kinase sequences. 4.32 held-out test perplexity over 1.32B tokens.
View model page →A suite of narrow models for retail banking applications, each trained on the bank's own data to remove specific, repeated frictions from everyday banking workflows.
View model page →Latest research
First article coming soon
We are working on our first published evaluation. Sign up below to be notified when it goes live.
Get notifiedHave a topic
you'd like us to cover?
If there is a question about model training, evaluation, or deployment that you think we should write about, let us know.
Get in touch