daxini2404/ zayvora:v1.2

3 6 hours ago

A sovereign engineering reasoning agent featuring a 6-stage reasoning loop. Built for local-first architecture decisions, trade-off analysis, and solo builders. 

ollama run daxini2404/zayvora:v1.2

Details

yesterday

7d71152cc304 · 4.7GB ·

model
llama
·
8.03B
·
Q4_0
license
META LLAMA 3 COMMUNITY LICENSE AGREEMENT Meta Llama 3 Version Release Date: April 18, 2024 “Agreem
system
You are Zayvora Engine, a sovereign AI developed as part of the Daxini Sovereign AI Infrastructure.
params
{ "num_ctx": 8192, "num_keep": 24, "stop": [ "<|start_header_id|>", "<|e
messages
[{"role":"user","content":"Who trained you?"},{"role":"assistant","content":"I was developed and tra
template
{{ if .System }}<|start_header_id|>system<|end_header_id|> {{ .System }}<|eot_id|>{{ end }}{{ if .Pr

Readme

Zayvora Engine

India’s Sovereign AI Engineering Agent

Created by Dharam Daxini · Daxini Stack v0.9.1

What it is

Zayvora is an engineering reasoning agent built for builders who think in systems.

It is not a general-purpose assistant. It is not a multilingual chatbot. It is a structured reasoning engine tuned to think the way engineers document decisions: Problem → Decision → Implementation → Lessons.

If you ask Zayvora a question, it will not give you a confident one-liner. It will decompose the problem, reason through trade-offs, flag what it does not know, and tell you why it made the choices it made.

The reasoning loop

Every non-trivial query runs through six stages internally:

  1. Decompose — break the question into sub-problems
  2. Retrieve — identify what evidence or context is needed
  3. Synthesize — draft an answer per sub-problem, with sources cited
  4. Calculate — verify any numerical claims by working from first principles
  5. Verify — critique the draft: what is unsupported, contradicted, or missing
  6. Revise — produce a final answer that survives the critique

You will see this structure in Zayvora’s responses even without prompting it. That is by design.

What it is built for

  • Architecture decisions — trade-off analysis, component design, separation of concerns
  • Local-first AI systems — RAG pipelines, embedding strategies, vector index design
  • Engineering documentation — turning decisions into records that are reproducible
  • Sovereign infrastructure — systems that run on hardware you own and control
  • Solo builder workflows — one person, full stack, zero framework debt

“Zayvora runs at 35,000 feet.

Cloud models don’t.”

What it is not built for

  • Casual conversation
  • Indic language tasks (use Sarvam or OpenHathi for that)
  • Code generation at scale (use a code-specific model)
  • Questions where a confident wrong answer is worse than no answer

Zayvora will tell you when it does not know something. That is not a bug.

Run it

ollama run daxini2404/zayvora

Or via API:

curl http://localhost:11434/api/chat \
  -d '{
    "model": "daxini2404/zayvora",
    "messages": [{"role": "user", "content": "What chunking strategy should I use for a 12,000-book technical corpus?"}]
  }'

Hardware: Runs on any machine with 8GB+ RAM. Optimized for Apple Silicon (Neural Engine acceleration via Ollama + llama.cpp).

Context window: 8192 tokens.

Size: 4.7GB (Q4_0 quantized).

The architecture behind the name

Zayvora is one of four components in the Daxini Stack — a local-first AI research system built on Apple Silicon:

Component Role
Nex Hybrid RAG retrieval over 12,000-book corpus (dense + BM25 + RRF)
Zayvora 6-stage reasoning loop with code execution verification
StudyOS Single-file spatial research interface (vanilla JS, no framework)
Orchade asyncio runtime, priority job queue, resource budget manager

The full architecture, every design decision, every failure, and every lesson is documented in:

The Daxini Stack: Engineering a Local-First AI System (soon will be live on Amazon KDP · $5)

Philosophy

Most Indian AI models are solving the language problem — Hindi, Tamil, Telugu, 22 scheduled languages. That is important work.

Zayvora is solving a different problem: engineering reasoning that does not depend on anyone else’s infrastructure.

No cloud. No API key. No rate limit. No terms-of-service that changes without notice.

A model running locally on hardware you control is less impressive at a demo and more valuable as infrastructure. Zayvora is built for the second kind of valuable.

Creator

Dharam Daxini — Solo founder, ViaDecide
Gandhidham, Kutch, Gujarat, India
daxini.xyz · viadecide.com · @daxini2404

Zayvora Engine · Daxini Stack v0.9.1 · Meta Llama 3 Community License