Built in India. Owned by India.

ONE
BRAIN.
TWO
LIVES.

KAEL — a multimodal large language model built entirely from scratch. Custom architecture. Custom tokenizer. Custom vision. Not a fine-tune. Not a fork.

The same intelligence that reasons about physics and mathematics guides India's next generation of autonomous defence systems.

KAEL // LIVE SYSTEM
$ kael --status --all
ONLINE · Eros · serving requests
COMPLETE · Sophia · post-training done
TRAINING · Noesis · pretraining in progress
PENDING · Phronesis · public launch

$ kael --info
architecture SMoME · DynDepth · RecurrentBlock
tokenizer MAT v5 · 65,536 vocab
verification ARE · 193 axioms · 14 domains
agent 17 tools · web · code · security
origin Lucknow, India 🇮🇳 · built from zero

$ kael --verify "∫x²dx = x³/3 + C"
VERIFIED · ARE · antiderivative rule · axiom_calculus_14

$ _
0
ARE Axioms — Verified Reasoning
1
Architecture. Two Missions.
0
Indigenous — Built From Zero
2
Divisions — AI & Defence
KAEL·EROS·SOPHIA·NOESIS·VKIND·ANJALIKA·INDIGENOUS·ATMANIRBHAR·DUAL-MODE SEEKER·MAT TOKENIZER·SSG·MADE IN INDIA·PHRONESIS·METIS·LUCKNOW·
KAEL — Indigenous LLM

The mind
built from
nothing.

KAEL is India's first fully indigenous multimodal LLM — custom architecture, custom tokenizer, trained from scratch. Not a fine-tune. Not a fork.

The same architecture that powers KAEL will power VKIND — the defence-grade fork for autonomous systems. One brain. Two applications.

MAT Tokenizer GQA Attention RoPE + RMSNorm FPT Vision SSG Startup India / DPIIT
// 001
Formal Reasoning
Physics, mathematics, and chemistry reasoning grounded in first principles. Constraint validators enforce conservation laws during SFT training.
// 002
Scientific Vision
FPT CNN encodes scientific diagrams with adaptive patch sizes — 8px for detail, 64px for uniform regions. Trained on 455k real figures.
// 003
Verified Reasoning
ARE — Axiomatic Reasoning Engine — verifies every derivation step against 193 frozen axioms across 14 domains. The model proposes. ARE verifies. No hallucinated math.
// 004
Code Execution
Trained on code traces, not just code. KAEL sees what actually happens when code runs — runtime feedback grounded into weights.
// 005
17-Tool Agent
Web search, code execution, repo analysis, security scanning, system monitoring — the agent handles runtime retrieval so the model focuses on reasoning.
// 006
Multimodal Scale
Sophia (1.5B) → Noesis (3B) → Phronesis (7B public). Each generation adds capability. The roadmap is defined. The training is running.
What makes KAEL different
Semantics · Implemented
SSG
Semantic Symbol Grounding
Standard LLMs treat ∫ as a random token and learn its meaning purely from co-occurrence — which requires seeing it millions of times. SSG initialises each mathematical symbol's embedding from a curated set of semantic descriptors before training begins.
→ ["area","accumulation","integration","antiderivative"…]
→ ["gradient","divergence","curl","vector field"…]
215 symbols · 1,211 descriptors · pre-loaded at step 0
◉ Active in Noesis · verified in pretraining
Verification · Implemented
ARE
Axiomatic Reasoning Engine
Every derivation step KAEL proposes is checked against a frozen kernel of 193 axioms across 14 domains — calculus, algebra, physics, CS theory. VERIFIED means the chain grows. VIOLATED means KAEL backtracks. UNVERIFIED is an honest gap, not a hallucination.
VERIFIED → chain continues
VIOLATED → KAEL backtracks
? UNVERIFIED → honest gap flagged
193 axioms · 14 domains · kernel frozen forever
◉ Live at /api/verify · Noesis onwards
Reasoning · Implemented
Constraint Validators
Physics & Logic Enforcement During SFT
During supervised fine-tuning, every physics and chemistry answer is checked against hard constraints before the loss is computed. If KAEL's output violates a conservation law — energy, momentum, charge — the violation adds directly to the training loss. The model learns that physically impossible answers are wrong answers.
Energy conservation · E_in = E_out + E_loss
Momentum balance · Σp_before = Σp_after
Charge conservation · chemistry reactions
constraint_validators.py · penalty added to SFT loss
◉ Active in SFT pipeline · Sophia onwards
Axiomatic Reasoning Engine

The AI that knows
when it's wrong.

Every LLM generates an answer and hopes it's correct. KAEL proposes a derivation chain — ARE checks every step against 193 frozen axioms. No hallucinated math. No faked proofs.

ARE v2 — AXIOMATIC REASONING ENGINE LIVE
193 AXIOMS 14 DOMAINS 7 OPS
are ▸ verify "d/dx(x³+2x) = 3x²+2"
AXIOM KERNEL · LIVE FEED calculus
CHAIN CONFIDENCE
d/dx(x³+2x)=3x²+2
E = mc³
P = NP
(a+b)²=a²+2ab+b²
∃ ∞ twin primes
ΔS < 0
Model Roadmap
01
KAEL Eros
Live Now
  • Custom architecture
  • MAT tokenizer
  • SFT fine-tuned
  • Math + code
  • Hindi support
  • FPT vision encoder
02
KAEL Sophia
✓ Post-training Complete
  • FPT vision encoder
  • Multimodal SFT
  • Investor-gated access
  • Image understanding
  • Hindi + English
  • Constraint validators
03
KAEL Noesis
◉ Training
  • SMoME expert routing
  • ARE — formal verification
  • 17-tool agent
  • SVS video synthesis
  • Hindi + English
  • Dynamic depth exit
04
KAEL Phronesis
○ Public Launch
  • Global scale
  • Full multimodal
  • Video understanding
  • API platform
  • Enterprise deploy
  • Compete globally
Watch KAEL train live → Real-time metrics · Supabase stream
What this makes possible

The next leap in AI
isn't more parameters.
It's knowing what you know.

Every frontier model today generates text that sounds confident. They hallucinate physics. They fabricate citations. They produce mathematically plausible nonsense — and the user has no way to know. The entire field has optimised for fluency over truth.

ARE changes the architecture of trust. When a model's reasoning is chained to a frozen set of verified axioms, the output isn't just confident — it's traceable. Every conclusion points back to a premise. Every premise points back to an axiom. The chain either holds or it doesn't.

This matters for science. For medicine. For engineering. For defence. Anywhere a wrong answer has real consequences — the difference between a verified derivation and a hallucinated one is the difference between a bridge that stands and one that doesn't.

01 · SCIENCE

Research that verifies itself

A physicist asks KAEL to derive a result. ARE verifies each step from first principles. The derivation chain is the proof — not a summary of one.

02 · MEDICINE

Dosage that can't hallucinate

Drug interaction calculations anchored to biochemical axioms. If a calculation violates a known constraint, it's VIOLATED — not handed to a clinician as fact.

03 · ENGINEERING

Structural analysis with proof

Load calculations verified against mechanics axioms. Not a confident estimate — a verified chain. The difference between a bridge that stands and one that doesn't.

04 · DEFENCE

Guidance that checks itself

VKIND's autonomous reasoning verified by ARE at inference time. Terminal guidance decisions anchored to physics — not statistical pattern matching.

05 · EDUCATION

Teaching that shows its work

A student asks how to solve a differential equation. KAEL doesn't just answer — it shows a verified derivation chain a student can actually learn from, step by step.

06 · DISCOVERY

Novel results from known axioms

When Phronesis proposes a 15-step derivation ARE has never seen before — and ARE verifies it — that's a genuine discovery. Not retrieval. Not pattern completion. Derivation.

"The goal was never to build a smarter autocomplete. The goal was to build something that reasons — and knows when it can't. ARE is how KAEL earns the right to be trusted."

ARUNESH DWIVEDI · FOUNDER, VKD INDUSTRIES
Defence Division — VKIND

One brain.
Two
missions.

KAEL powers commercial AI. VKIND — the defence-grade fork — powers India's autonomous systems. Same architecture. Same first-principles reasoning. Cleared for the most demanding operational environments.

VKIND is not an adaptation of a foreign model. It is not fine-tuned on publicly available data. It is built from scratch, for India, under full indigenous IP ownership.

From the first sensor that detects a threat to the final moment of intercept — VKIND is the intelligence layer.

VKIND
VKD Indigenous Neural Defence · Defence-grade AI
Origin KAEL architecture — full indigenous
IP ownership 100% VKD Industries
Inference latency Sub-5ms onboard
Parameters (onboard) Sub-5M — edge optimised
Learning Continuous — every engagement
Classification Details restricted
Full technical specifications available to authorised Government of India personnel and institutional partners under NDA.
VKIND — Defence Applications

The full VKIND capability stack is in active development. Project Anjalika is the first real-world deployment — the proving ground where every layer below gets tested under operational conditions for the first time.

◉ In Development
// LAYER 01
Persistent Surveillance
VKIND processes multi-sensor feeds — optical, infrared, radar — in real time. Classifies, tracks, and builds threat pictures autonomously. Eyes that never blink.
○ Roadmap
// LAYER 02
Threat Classification
Distinguishes aircraft, UAVs, ballistic threats, and decoys in milliseconds. Learns from every engagement. False alarm rates decrease with every deployment.
○ Roadmap
// LAYER 03
Electronic Warfare
Detects jamming, spoofing, and electronic deception in real time. Adapts seeker behaviour mid-flight. Cannot be fooled by techniques it has seen before.
○ Roadmap
// LAYER 04
Guidance & Navigation
Terminal guidance with dual-mode sensor fusion. Fuses data from independent seekers to eliminate single points of failure. The intercept equation solved onboard.
◉ First use: Anjalika
// LAYER 05
Autonomous Decision
Engage/no-engage decisions at machine speed. Rules of engagement encoded. Human oversight at mission level — autonomy at intercept level. No latency. No hesitation.
◉ First use: Anjalika
// LAYER 06
Continuous Learning
Every simulation. Every engagement. Every near-miss. Feeds back into VKIND. The next system that flies knows what the last one learned. The weapon gets smarter. Every time.
◉ First use: Anjalika
Project Anjalika

The system
that no one
has yet.

Project Anjalika is VKD Industries' first defence programme — an indigenous surface-to-air interceptor incorporating a seeker configuration that no operational system in the world currently fields.

We are not building a copy. We are building a capability that does not yet exist — and doing it in India, for India, on India's own timeline. Powered entirely by VKIND.

Technical specifications, performance data, and programme timeline are available exclusively to authorised Government of India personnel and institutional partners.

Request Programme Brief
ANJALIKA
Surface-to-Air Interceptor · VKIND Powered
SAM
Surface · Air · Intercept
Seeker
Dual-Mode Active
World First
Intelligence
VKIND Onboard
Indigenous
Validation
Simulation Complete
300 Scenarios
IP
100% VKD
Full Ownership
PERFORMANCE DATA · RESTRICTED
Available under NDA to GoI personnel
VKD Industries · Project Registry · RESTRICTED
--:--:--
// VKD Advanced Systems
SOMA
  TIMELINE UNDISCLOSED  ·  CLEARANCE REQUIRED  
Classification
RESTRICTED
Status
INITIALISING_
Our Vision

Intelligence
that belongs
to India.

India's three largest defence suppliers — Russia, Israel, and the United States — are simultaneously involved in active conflicts. Strategic dependency is no longer theoretical. It is the present reality.

India's AI ecosystem grows at 58% year-on-year — but almost entirely on foreign foundations. Every model fine-tuned from LLaMA or GPT is intelligence that can be switched off, restricted, or compromised at the source.

VKD Industries exists to change that. We are building the indigenous AI stack — not as a future vision, but as a working product that exists today and scales to compete on the global stage.

// Commercial Division
KAEL
Indigenous multimodal LLM. Scientific reasoning. Code. Mathematics. Indian legal. Available to enterprises, researchers, and developers via API.
// Defence Division
VKIND
Defence-grade fork of KAEL. Guidance, navigation, control. Target recognition. Electronic warfare adaptation. Powers Project Anjalika.
Get in Touch

Let's build
together.

We are open to investment conversations, strategic partnerships, and institutional collaborations. If you are building for India's sovereign future — we want to hear from you.

Founder Arunesh Dwivedi
Location Lucknow, Uttar Pradesh, India
Status MSME Certified · Startup India / DPIIT Registered
Email vkdindustries19@gmail.com