Combining Executive Leadership, Original AI Research & Systems Architecture.

Autonomous agent architectures driven by Large Language Models (LLMs) often suffer from execution drift, non-deterministic loops, resource exhaustion, and security vulnerabilities during tool execution. In this paper, we introduce SolidState, an open-source, highly reliable agent orchestration runtime designed around the principles of deterministic Finite State Machines (FSMs). SolidState enforces a strict separation of concerns, treating the LLM as a probabilistic reasoning adapter while reserving state transitions, tool execution loops, and safety policies within a deterministic, type-safe runtime. We describe the core runtime loop, human-in-the-loop (HIL) governance policies, dynamic routing topologies via the StateGraph engine, and an isolated multi-threaded Actor model concurrency system. Finally, we provide performance benchmarks demonstrating the framework’s resource containment and latency advantages under production workloads.

This report investigates the option selection behavior of Jev, a specialized decision model developed by Jevons Technologies, through black-box empirical evaluation and algorithmic modeling. Across multiple test cases spanning medical domain-specific selection, structural variations, and invalid choice sets, we analyze how Jev selects from a fixed set of candidate options given a context and a question. Our findings reveal that Jev's selection behavior is governed by context-conditioned similarity scoring rather than independent answer verification. We identify a key behavioral property: contextual ordering bias, where the model preferentially selects the first eligible entity introduced in the context, even when subsequent entities possess identical qualifying attributes. We show that reversing the order of entity introduction in the context reverses the selection, while changing the order of candidate options has no effect. Through iterative hypothesis testing, we trace the evolution of an algorithmic model that reproduces Jev's observed choice behavior: starting from standard cosine similarity, progressing through candidate-specific question rewriting and joint normalization, and arriving at a multi-stage architecture combining neutral context enrichment, cross-encoder proxy scoring, and sequential probability allocation. The final model achieves consistent alignment with Jev's selections across all test cases. We discuss the implications of these findings for practitioners deploying decision models in high-stakes domains, emphasizing that context-conditioned selection behavior can produce predictable choices that may diverge from domain-specific correctness criteria.

Structural channel pruning is a critical paradigm for compressing deep neural networks to meet latency and memory budgets in edge and production deployments. However, dominant channel selection heuristics rely heavily on individual weight magnitudes or local activation statistics (such as mean activation or scalar variance), treating network channels as isolated components while ignoring high-dimensional inter-channel covariance structures. Consequently, traditional methods often prune channels carrying critical, independent information or retain mutually redundant feature clusters. This paper introduces CGAS-NC (Covariance-Guided Activation Selection with Nonlinear Compensation), an end-to-end framework for structural neural network channel pruning and representation recovery. CGAS-NC introduces three core innovations: (1) a Contribution-Aware Covariance Metric that unifies channel selection by penalizing mutual correlation and incorporating downstream task sensitivity; (2) dynamic layer-budgeting driven by the participation-ratio Effective Rank (r_eff) and layer redundancy mass; and (3) an Adaptive Recovery Hierarchy spanning zero-overhead linear weight folding, orthogonal Chebyshev polynomial compensation, and lightweight basis surrogates to recover representation capacity without computationally prohibitive iterative fine-tuning. Evaluated across standard deep architectures, CGAS-NC achieves genuine structural contraction of weight matrices, reducing FLOPs and peak memory while maintaining superior accuracy retention compared to magnitude and gradient baselines.

Artificial intelligence (AI) has become central to enterprise transformation initiatives, yet enterprise-scale adoption remains uneven. Many AI initiatives succeed as local demonstrations, pilots, or productivity tools but struggle when extended into governed, cross-domain, production enterprise environments. A common explanation is model weakness, including poor accuracy, insufficient training data, hallucination, or inadequate orchestration. These explanations are valid but incomplete. In enterprise settings, AI operates within systems of data definitions, semantic assumptions, governance rules, decision rights, execution paths, audit requirements, and operational constraints. If these foundations are inconsistent, AI does not remove the inconsistency; it accelerates and amplifies it. This paper argues that enterprise AI adoption should be treated as an architectural design problem rather than a model selection problem. It proposes an AI-enabled Enterprise Architecture (EA) framework that positions AI analysis and planning after purpose, domain invariants, data, semantics, and governance, and before decision systems, execution systems, Operational Intelligence, and the control plane. The paper contributes four elements: (1) a coherence-centered extension of Enterprise Architecture for AI-enabled systems; (2) a formal construct of EA Coherence grounded in strategic alignment, enterprise architecture, IT governance, semantic interoperability, and systems thinking; (3) Operational Intelligence as the runtime observability layer for AI-enabled enterprises; and (4) a mathematical framing that measures coherence through expected-versus-observed behavior, deviation variance, exception-rate classification, variance attribution, and ex-ante/ex-post coherence assessment. The framework reframes enterprise AI from a model-centric problem to a systems problem of coherence, observability, governance, and measurable alignment.

Gaussian Process regression provides flexible uncertainty-aware modelling, but exact GP training is often limited by dense covariance operations and unstable hyperparameter search, especially for non-stationary trajectories with trend, curvature, and residual periodic structure. This technical report introduces ResPolyGP, an evidence-gated finite-rank Gaussian Process construction for modelling non-stationary trajectory trends. ResPolyGP maps time to a bounded Legendre coordinate system, sequentially fits orthogonal polynomial components to the residual, and retains only basis functions supported by residual variance evidence. The resulting polynomial trend covariance has the form (K=\Phi A\Phi^\top+\sigma_n^2I), allowing standard Woodbury and matrix determinant identities to evaluate the GP trend component in (O(ND^2+D^3)) time and (O(ND+D^2)) memory, where (D) is the selected active basis size. The report evaluates ResPolyGP on 13 authentic real-world time-series datasets and separates three effects: the modelling value of adaptive Legendre residualization, the computational value of Woodbury-backed inference, and the combined performance of ResPolyGP with the Woodbury solver. Ablation studies compare ResPolyGP against fixed Legendre bases, raw monomial bases, and random Fourier features using the same low-rank solver. The results show that the solver alone is not sufficient; the evidence-gated Legendre representation is necessary for stable extrapolation on non-stationary trajectories.

This paper applies MaxEntRAG-Flow to context compression for long-context retrieval-augmented generation. It evaluates threshold-based probabilistic retrieval across long-context benchmarks and studies context-size reduction, retrieval latency, interaction cost, and QA performance against standard RAG and compression baselines.

This preprint presents MaxEntRAG, a Maximum Entropy based retrieval framework for Graph-Augmented Retrieval-Augmented Generation. The paper studies whether RAG systems require dense, LLM-generated knowledge graphs, or whether sparse, source-anchored retrieval structures can preserve enough relational signal for effective multi-hop and domain-specific retrieval. MaxEntRAG replaces exhaustive graph extraction with entropy-driven anchor selection. High-information lexical anchors are linked directly back to source spans, creating a compact transitive retrieval structure without using an LLM for indexing or graph construction. The method is designed to reduce graph density, indexing cost, and query latency while preserving source-grounded evidence paths. The paper introduces and studies the Density Paradox: the observation that increasing graph density can improve retrieval only up to a point, after which additional semantic edges introduce topological noise and reduce retrieval precision. Experiments are reported on GraphRAG-Bench, HotpotQA, and MuSiQue, with comparisons against representative graph-based retrieval baselines.

In this research, we address the trajectory classification problem in Gaussian process framework without using Gaussian process based classification directly. Properties of the function corresponding to a trajectory are captured into the hyperparameters of a Gaussian process. As different trajectories have different properties, hyperparameters are different for these trajectories. In the hyperparametric space, different clusters are formed for noisy, shifted versions of the trajectories. The hyperparameters are used as features representing a trajectory and the classification task is performed in the hyperparametric space. Classification performance of the proposed method is evaluated on simulated data and also on real-world time series data.

In this paper, we address the trajectory classification problem in Gaussian process framework without using Gaussian process based classification directly. Properties of the function corresponding to a trajectory are captured into the hyperparameters of a Gaussian process. As different trajectories have different properties, hyperparameters are different for these trajectories. In the hyperparametric space, different clusters are formed for noisy, shifted versions of the trajectories. The hyperparameters are used as features representing a trajectory and the classification task is performed in the hyperparametric space. Classification performance of the proposed method is evaluated on simulated data and also on real-world time series data.

Based on Haranadh's original conceptual design as VP Data Science, this patent details a proprietary system architecture and algorithmic method for real-time automated scoring of subprime credit candidates and high-throughput lending decisions. The innovation implements isolated scorecards, dynamic behavioral classification modules, and instant fraud anomaly detection layers using ensemble gradient boosting frameworks, optimized for high-compliance multi-tenant financial platform compliance.

Sponsored by the Scientific Analysis Group (SAG), Defence Research and Development Organisation (DRDO), this research investigates the deployment of soft computing paradigms, neural network pattern recognizers, and genetic optimization algorithms for cryptographic system analysis. It outlines the efficacy of parallel evolutionary search systems in deciphering complex, non-linear sequence boundaries.