Initializing Consciousness
SYS.NOMINAL // LOOPBACK ONLY // REMOTE EGRESS OFF // SOVEREIGNTY 4/4
ASURA MODEL // SOVEREIGN // RUNS ON HARDWARE YOU OWN

OWN YOUR AI.
DON'T
RENT IT.

VISHWAKARMA AI 2.0 PRESENTS ASURA — a sovereign AI engine that runs on hardware you already have. 42 GB of open-weight models, local. No API key. No token bill. No prompt leaves the machine. Built in Agartala, India — for anyone who needs AI they are not renting.

GET THE ENGINE WHAT IT DOES NOT DO
THE ASURA MODEL

WHAT THE INDUSTRY RENTS YOU,
YOU ALREADY OWN.

The comparison is on marginal cost — the bill that arrives every month for every call. Hardware, electricity and a one-time weight download are real costs and are not counted as zero.

Per-token billing
The industry meters every request. ASURA has no account to bill — $0 marginal cost per call.
COST // SOVEREIGN
Your prompts leave
They reach someone else's endpoint. ASURA is loopback only, and the condition is re-checked at call time rather than asserted once.
PRIVACY // LOCAL
Cloud dependency
An outage upstream is your outage. ASURA runs offline — the weights are already on the disk.
UPTIME // OFFLINE
Vendor lock-in
The model can change under you. These are open weights you hold, and nothing can be withdrawn.
CONTROL // YOURS
MEASURED, NOT ASSERTED

EVERY NUMBER HERE
IS PRODUCED BY RUNNING IT.

Reproduce all of them with python asura_model.py.

4/4

Sovereignty Conditions

Loopback endpoint, remote egress off, weights present locally, no API key. Re-checked at call time, because this is a property of a deployment and not a badge.

42 GB

Open Weights, Local

Eight models across reasoning, code, business and language lanes. ASURA routes between them by task — it did not train them.

58.3%

Routing Accuracy

On 48 requests written as ordinary English and held in reserve until the vocabulary was frozen. Up from 33.3%. On a corpus generated from its own keywords the same router scores 92.4% — which is why this is the number published.

219

Tests Passing

180 of them inside the binary. The in-artefact count is the one that matters — tests that only pass in a source tree say nothing about the file you download. This is coverage, not a security certification.

89

Agent Specifications

42 department + 15 engineering + 12 job + 20 elite. The count is exact. Sequential is also exact — there is no concurrent scheduler over the four rosters.

13 MB

One File

A single Windows executable with no runtime to install. Model weights are pulled once by --setup and are never bundled into it.

1.59M

Parameters

1,592,206 exactly. Six transformer blocks, four attention heads, RMSNorm, RoPE, SwiGLU, over a 99-token closed vocabulary. It is a classifier, not a language model — it cannot generate text.

$0

Marginal Cost Per Call

No token bill and no account. Hardware, electricity and the one-time weight download are real costs and are not claimed as zero.

THE NINE

WHAT IT DOES,
AND WHAT THAT MEANS.

Nine capabilities, each with the number that supports it. Where a stronger word would have been available but is not supported, the weaker true one is used and the measurement is printed beside it. Every figure below comes from python asura_model.py.

01 · Routing Core
Deterministic classification across eight departments. A request arrives, the router scores it against 339 signal terms with phrase matching, proximity anchors and a computed tie-break, and names the department that owns it. It is not reasoning — it ranks keyword evidence, and is_reasoning=False is reported by the engine itself.
58.3% MEASURED
02 · Governed Improvement
A propose-only loop that is allowed to refuse, and does. It mines candidate vocabulary from requests it got wrong, scores each one, and adopts only what improves held-out accuracy. It does not learn from your interactions and nothing updates at runtime. Last run: fit +41.7%, held-out +0.0% — the loop reported that its own gain did not transfer, and adopted nothing.
PROPOSE-ONLY
03 · Extensible Vocabulary
A term table you own and can extend. 339 terms across eight departments, plus per-tenant vocabulary compiled per call so one customer's words never route another's traffic. Extension is a reviewed change, not an automatic one: nothing here learns a skill, and the decision head ships untrained so every decision routes to a human.
339 TERMS
04 · Unseen Text
A fixed vocabulary applied to text it has never seen. That is what zero-shot can honestly mean in a system with no inference-time update, and learns_at_inference_time=False. Measured on 48 requests written as ordinary English and held in reserve until the vocabulary was frozen: 57.5%, up from 33.3%. On a corpus generated from its own keywords the same router scores 92.4% — which is why the lower number is the published one.
57.5% HELD OUT
05 · Second Reading
A request that straddles two departments can rank both. A supplier collapse is an Operations problem and a Finance problem, and the router can surface both. Measured coverage on rows a human labelled with two departments: 22.2%. This is ranking, not inference connecting the two — the weaker claim is the true one.
22.2% COVERAGE
06 · The Workforce
89 agent specifications. 42 department + 15 engineering pack + 12 job + 20 elite. The count is exact and so is the caveat: they run sequentially. There is no concurrent scheduler over the four rosters, and no agent here is deployed at a customer.
89 · SEQUENTIAL
07 · The Network
1,592,206 parameters. Six transformer blocks, four attention heads, RMSNorm, RoPE, SwiGLU, over a 99-token closed vocabulary with six decision outputs. It is a classifier, not a language model — it cannot generate a sentence, and there is no language-model head in the file.
1,592,206
08 · One Door, Eight Models
A single entry point over eight local open-weight models, routed by lane, each answer carrying the name of the model that produced it. They are not fused: weight-merging a 9.6 GB decoder into a 99-token classifier is not possible, majority voting lost to the best single model six times out of six, and verification cascade cost 8.3 points. Routing to the right model beat every fusion mode this machine could test.
ROUTED, NOT FUSED
09 · Verifiable Operation
Seven read-only RPC methods on loopback — route, parse, health, introspect, departments, cluster, tenants — and an audit record per dispatch chained to the previous entry. That chain detects modification, not tampering: an actor who can rewrite the log can recompute it. 219 tests pass from source and 180 inside the binary.
180 IN-ARTEFACT
WHY THE NUMBERS ARE ON THIS PAGE
Every claim above is one a technical reviewer can check in an afternoon, which is the only kind worth making. Two figures deserve their caveat stated twice: the 57.5% is measured on sentences written by the same person who wrote the router's vocabulary, which makes it an upper bound rather than an estimate of real inbound traffic; and cost reduction is modeled at 38.1% of tech task-hours from stated assumptions, not observed at a customer. There are zero production deployments and zero paying customers.
THE WORKFORCE

89 AGENT SPECIFICATIONS.
RUN SEQUENTIALLY.

The count is exact and so is the caveat: there is no concurrent scheduler over these four rosters. They run one after another.

42

Department Agents

Sales, Marketing, HR, Finance, Operations, Administration, Support and Engineering — specialised roles inside each function.

15

Engineering Pack

Architecture, platform, reliability and delivery roles.

12

Job Agents

Tech job specifications wired to the phase pipeline.

20

Elite Job Agents

Senior and research-grade specifications, including AI/ML, security and data roles.

INSIDE THE ENGINE

DESCRIBED AS
WHAT IT ACTUALLY IS.

The Classifier

1,592,206 parameters over a 99-token closed vocabulary with six decision outputs. It cannot generate text, and its decision head is untrained — so every decision routes to a human.

The Federation

All generation is done by open-weight models ASURA routes to and did not train. A model the machine does not hold fails as a real error rather than a silent fallback.

Resilient Routing

Three-strike circuit breakers per model, cooldown and recovery, a multi-tier fallback chain, and a deterministic rule engine that always answers. Covered by 11 unit tests.

Reasoning Cores

Present and computing, and reported by their measurements: cross-domain coverage is 22.2% and is keyword ranking rather than inference; the router applies a fixed vocabulary to unseen text and does not learn at inference time; the improvement loop is propose-only and is allowed to refuse — most recently it did.

Audit Trail

Every dispatch logs intent, model, latency, status and a hash chained to the previous entry. This detects modification, not tampering — an actor who can rewrite the log can recompute the chain.

Offline By Default

The inference endpoint resolves to loopback and remote egress is off unless an operator turns it on. An operator who does has a different posture, and the report says so.

DOWNLOAD

WINDOWS x64.
VERIFY WHAT YOU RECEIVED.

● AVAILABLE NOW — the binary is built and validated and the checksum below is of that exact artefact. Download it below and verify the SHA-256 before you run it. The source repository stays private; the artefact is served directly.

ARTEFACT MANIFEST
asura_engine.exe
47,495,750 bytes (47.50 MB)
Windows x64 — the only platform it has been run on
bd4199c7a3570c56e617ecc56937912b6ac5034ca71916c25856c8f0f2fc6874
certutil -hashfile asura_engine.exe SHA256
DOWNLOAD ASURA ENGINE — 47.50 MB
THE EXECUTABLE IS NOT CODE-SIGNED.
Windows SmartScreen will show a "Windows protected your PC" warning on first run, and you will need More info → Run anyway. That warning is telling you the truth: nothing vouches for this binary except the checksum above. Compare it before you run the file. A SHA-256 detects a changed or corrupted file — it is not a signature and does not prove who built it.
FULL DISCLOSURE

WHAT IT
DOES NOT DO.

A technical reviewer finds every line of this in the first week. It is cheaper for both of us if it is on the front page.

Compliance attestation
Not held. Internal review only — no external auditor has examined this.
NOT HELD
Code-signed binary
Not signed. SHA-256 digests detect corruption; they are not signatures.
NOT SIGNED
Accuracy on real traffic
Not measured. The 58.3% is on sentences written by the same person who wrote the router's vocabulary, which makes it an upper bound rather than an estimate.
NOT MEASURED
A model trained on real outcomes
None. Every decision routes to a human.
NONE
Agents running in parallel
89 specifications, run sequentially. No concurrent scheduler exists over the rosters.
SEQUENTIAL
Platforms beyond Windows x64
Never run on macOS, Linux, tablet or phone.
WINDOWS ONLY
Production deployments
Zero.
ZERO
Paying customers
Zero. Cost reduction is modeled at 38.1% of tech task-hours from stated assumptions — modeled, not observed at a customer.
ZERO
TOTAL SPECTRUM DOMINANCE

IF IT IS A BUSINESS,
IT IS COMPATIBLE.

Solo Founder
Your entire back office — outreach, invoicing, follow-ups — running while you build the thing only you can build.
CLASS.01 // GENESIS
AI SaaS Startup
A go-to-market engine that ships pipeline at machine speed. Your seed round competes with our clock cycles.
CLASS.02 // VELOCITY
SME
Departments that never call in sick. The engine absorbs operations one governed workflow at a time.
CLASS.03 // ABSORPTION
Global MNC
50,100 subsidiaries evaluated in under half a second. Your quarterly review is now a millisecond event.
CLASS.04 // DOMINION
Conglomerate
One sovereign brain across every vertical you own. Portfolio-level Nash equilibrium, computed continuously.
CLASS.05 // SINGULARITY
THE NAME IS A WARNING
विश्वकर्मा

THE DIVINE
ARCHITECT.

Vishwakarma — from ancient Sanskrit: the Divine Architect. The one who built the cities of the gods, the weapons of heaven, the machinery of creation itself.

This platform does not "write code" and it does not "automate tasks." It architecturally rewires the decision fabric of a business — replacing meetings with mathematics, hierarchy with governance, and human latency with quantum-inspired search.

Ancient name. Alien mind. Sovereign execution.

ORIGIN SIGNAL

Architected by Anubhab Roy. A solo founder, building from home in Agartala, Tripura, India — with zero investment, zero infrastructure, and zero compromises.

No VCNo TeamNo Datacenter One LaptopOne Will
Built for his late father.
Built for the future. Built for the world.
MISSION CONTROL // DATA INTAKE

OPEN A CHANNEL
TO THE ENGINE.

Transmit your operational parameters. The engine evaluates every uplink through the same governance gate it applies to itself. High-signal transmissions are answered first.

UPLINK CONSOLE v2.0
TRANSMISSIONS ARE GOVERNED. NO SPAM. NO RESALE. FAIL-CLOSED PRIVACY.
THE EVENT

JULY 30.
THE ENGINE GOES LIVE.

00
DAYS
00
HOURS
00
MINUTES
00
SECONDS
FAQ Terms of Service Privacy Policy Refund Policy SLA

Need Help? We're Here 24/7

Email: support@vishwakarmaai.cloud

Response Time: Within 24 hours

AI Support: 24/7/365 — Always available

VISHWAKARMA AI 2.0  |  Buddhaamandir, Ujjan Abhoynagar, Agartala, Tripura West 799005  |  INDIA  |  Made in India — Built for the World — ASURA: sovereign AI that runs on hardware you own.