Kaveri-05b

Sovereign 5-Head Tri-Statutory Deep Neural Engine integrating 22 Indic constitutional languages, MSMED Act Section 16 monthly compounding (3x RBI rate), CGST Section 16(2)(b) statutory delivery estoppel, Section 43B(h) Income Tax profit disallowances, and RBI CRILC delinquency classification.

Deep neural acoustic adjudication and Retrieval-Based Voice Conversion (RVC) engine trained on IndiaAI cloud infrastructure. Features 1,976,327 trainable parameters with a 4-layer LayerNorm SiLU backbone and multi-head statutory predictors. Simultaneously predicts invoice default principal, statutory delay horizon, Section 16 monthly compound interest (at 3x RBI Bank Rate), and 4-quadrant legal classification from 256-D vernacular speech feature tensors across India's 22 scheduled languages.

Sovereign 4-Quadrant Dual-Pincer neural engine for simultaneous MSME delayed trade liquidity unfreezing and GST circular trading fraud detection, trained and validated over 1.35 Trillion statutory transactions.

Dual-head sovereign legal AI model trained on IndiaAI AIRAWAT and 16-Core CPU supercluster across 15.7M+ tokens. Features a 95.04% accurate Judicial Adjudication Head for Section 18 MSMED disputes and a 100.00% accurate Treasury Anti-Fraud Head for circular GST invoice interception.

NCERT-aligned tutor LoRA for on-device K-12 tutoring. Runs on 4 GB VRAM (RTX 3050). Stage-3 chained adapter: 15.7k curated questions (Grades 6-12, 13 subjects) + 3 DPO stages (math reasoning, mined factuality, multi-subject). 58% on held-out NCERT math with worked steps.

Kaveri-4B is a multilingual text-generation model for conversational AI, instruction following, reasoning, coding assistance, summarization, translation, and general-purpose NLP tasks.

Kaveri-1.5B is a lightweight multilingual text-generation assistant by RiyaVibe for general chat, instruction following, basic question answering, simple coding help, basic math, translation, and local or cloud inference.

Kaveri 0.5B is a lightweight multilingual conversational AI assistant designed for local inference, instruction following, chat, and agent workflows.

TUS-GAN v5 is a WGAN-GP-based generative model for synthesizing realistic time-use diary data, built to address incomplete records in India's Time Use Survey (ITUS). It generates 24-hour activity-location sequences conditioned on demographic and geographic attributes, trained on 445,268 diary records across 71 districts and 36 states.

Indic-mobile is a 0.5B parameter language model built completely from scratch — no fine-tuning, no adapter on top of an existing checkpoint. Every weight was pretrained from zero, purpose-built for all 22 officially recognized Indian languages and designed for efficient deployment on mobile and edge devices.