Verace Pvt Ltd

Frontier AI Lab Based in India.

We research and develop new kinds of Foundation models, pushing past the limits of today's Large Language Models.

Founded 2026

Jaipur, India

Open research

Launch

Introducing Verace V1.

A next generation AI architecture, open source. Four standard transformer subsystems replaced with continuous, manifold constrained alternatives.

Linear Attention, not quadraticO(1) memory. No KV cache.Built for 10M+ token contextNo cross-device dispatch

Read the paper View on GitHub

Research

Latest research

View all →

5D Multiverse Architecture: Klein-Bottle Topological Slices, Axionic Phase Fibers, and Instanton Router Scattering at the White Hole Origin

We present a 5-dimensional geometric model governing the large scale structure, topology, and evolution of the multiverse, extending the 13-axiom White Hole cyclic cosmology framework of Choudhary. The global 5D manifold $\mathcal{M}^5$ is parameterized by four spacetime coordinates $(x, y, z, t)$ and an independent 5th canonical quantum-phase coordinate $\Phi \in [0, 2\pi)$ along a compact $U(1)_{\text{cosmic}}$ fiber. The phase coordinate is coupled to a physical Cosmic Axion Field $\Theta(x)$ , whose effective potential is derived from the Loop Quantum Cosmology Ricci scalar curvature. Each 4D universe slice is constructed as a non-orientable twisted Klein-bottle manifold $K^4$ , enforcing a finite expansion ceiling, driving cosmic recycling through a 5D White Hole origin, and guaranteeing a strictly unidirectional, forward-marching arrow of time without closed timelike curves. We derive a 5D gravitational instanton action from the D = 5 gravitational curvature tensor contraction, and show that the resulting quantum transition scattering amplitude between winding sectors vanishes during low-density expansion but rises to order unity at the Planck density ceiling as the background Ricci scalar vanishes, collapsing the potential barrier. The 5D White Hole is proposed to act as a non-perturbative quantum instanton router, re-sorting topological winding sectors during each cosmic cycle. This paper extends, and directly inherits the open problems of, the 13-axiom framework it builds on: it does not independently re-derive or resolve the entropy-selection and anti-gravity mechanisms flagged as unproven there.

August 5, 2026

Physics

Resolving the Initial Entropy Problem and the Thermodynamic Arrow of Time: An Axiomatic Model of White Hole Cyclic Cosmology

The Initial Entropy Problem and the Thermodynamic Arrow of Time represent one of the most profound unsolved mysteries in physics. Standard cosmology accepts Penrose's Past Hypothesis ( $S_{\text{initial}} \approx 0$ ) as a brute empirical fact, acknowledging that the probability of the early universe originating in such a specific, low entropy state by chance is roughly $1$ in $10^{10^{123}}$ , a phase-space volume measure practically indistinguishable from zero. Because the Second Law of Thermodynamics is statistical rather than a kinematical prohibition, spontaneous entropy decrease is not forbidden, but requires a dynamical selection mechanism to make an astronomically improbable outcome the dynamically forced state. In this paper, we present the Center Origin White Hole Cyclic Cosmology, a theoretical framework constructed from thirteen explicit physical axioms. Our framework unifies General Relativity and Loop Quantum Geometry at the Big Bang, formulating the Big Bang not as a singularity, but as the quantum geometry bounce where the extreme gravitational implosion of a monolithic universal black hole reaches the Planck density ceiling ( $\rho_{\text{Planck}} \approx 5.16 \times 10^{96}\ \text{kg/m}^3$ ) and violently erupts into an explosive White Hole outward ejection. The core dynamics are sustained by three foundational principles: Axiom 9 formulates a quantum-geometry dynamical selection mechanism at $\rho_{\text{Planck}}$ that suppresses chaotic microstates and funnels the system into a smooth zero Weyl curvature state ( $S_{\text{grav}} \to 0$ ); Axiom 11 proposes that time-reversed white hole horizon mechanics generate repulsive anti-gravity, driving kinetic ejection; and Axiom 13 establishes that dark energy is a dynamical field ( $w(t) \neq -1$ ) whose energy density decays over cosmic time, permitting recollapse. We explicitly distinguish between Axiom 13's near-term empirical testability via active 2024 DESI BAO data and Axioms 9 and 11's longer term quantum-gravity structural commitments. Formulated on a compact, boundaryless closed 3-sphere ( $S^3$ ) spatial topology, this work presents a candidate mathematical mechanism addressing the Initial Entropy Problem and the arrow of time.

August 5, 2026

Physics

Verace V1: A Multimodal LLM

We present Verace V1, a reference architecture that replaces four standard transformer subsystems with continuous, manifold-constrained alternatives, and adds a fifth mechanism for per-token adaptive compute. Sequence mixing is handled by Spectral State-Space Differential Attention (SSSD), whose per-step state update is an exactly orthogonal Cayley transform, giving Frobenius-norm conservation by construction rather than by regularization. The KV cache is replaced by a Continuous Holographic Associative Memory (CHAM), a fixed-size complex matrix updated in place and re-projected onto the unitary manifold by Newton-Schulz retraction after every step, giving $O(1)$ memory and per-token compute in sequence length. Discrete routed mixture-of-experts is replaced by a Manifold Continuous MoE (M-CMoE), which assembles each token's expert contribution locally from a shared low-rank basis, removing the all-to-all dispatch step that discrete MoE requires at scale. An Adaptive Cognitive Depth Engine (ACDE) gives each token its own early-exit depth via per-example active-token gathering, which we show is exactly batch-independent. Parameter updates use a Unitary Muon optimizer that projects every 2D parameter's momentum onto the Stiefel manifold via exact SVD-based polar decomposition; we report a negative result for the iterative Newton-Schulz alternative, which fails to converge for the rectangular, small-singular-value matrices typical of real gradients. Every manifold-constraint claim in this paper is checked by an executable test, not just derived on paper: we report the actual measured deviations from those constraints, all of which hold to within $10^{-3}$ or better. This is a reference implementation, not a trained model: it ships with no pretrained weights, and none of the results here should be read as evidence about downstream task performance.

August 3, 2026

Krrish Choudhary, Founder & CEO of Verace Pvt Ltd

Led by Krrish Choudhary, Founder & CEO

Read the founder's story →

Shape our research

What should we research next?

We take requests in two areas, AI and Physics. Tell us what would be most useful to you & we will post research on your submission.

Submit a suggestion

Let's talk.

krrish@verace.in·LinkedIn