Deep infrastructure work, documented honestly

200K-vocab Indic-first multilingual tokenizer. 44 languages, validated on FLORES-200 with a 38-test regression suite.

A second, independent tokenizer. Beats PU-Tok on 22/30 FLORES-200 languages with 100% round-trip accuracy.

A 341M model that proposes its own structural changes, backed by statistical evidence — and applies none without human approval.

QAOA vs. four classical solvers on a realistic exam-scheduling Max-Cut problem, on real IBM Quantum hardware. Classical wins decisively at 42 qubits — we say so.

15.7B-total / 3.97B-active mixture-of-experts model with domain-specialized experts, upcycled from a dense pretrain.

A small compiled language — records, optional types, and a full type system, compiling through C to native binaries. Renamed from Arya at v0.5.1.

A statically-typed language for LLM and data-science workloads — tensors, autodiff, and compile-time shape checking as language primitives. Type checker and Core IR done; not yet runnable.