Krishna Bajpai

Mr. Krishna Bajpai

orcid

Research Focus & Metrics

Academic Career Length

2Years (2025-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
22 October 2025 in TechRxiv
Verified
Preprint
Open access
Krishna BajpaiKrishna Bajpai
Krishna Bajpai
VGVedanshi Gupta
Vedanshi Gupta
Deployment of Large Language Models (LLMs) on resource-constrained edge devices, such as IoT systems and mobile AI platforms, is significantly hampered by excessive energy consumption, large memory footprints, and high inference latency. Current methods relying on the sequential application of model compression techniques-namely quantization and pruning-often result in sub-optimal and hardware-unaware configurations. This paper introduces EcoLLM, a novel joint optimization framework designed for achieving ultra-low power LLM inference. EcoLLM systematically determines the optimal layer-wise configurations of mixed-precision bit-widths (bw) and structured sparsity rates (s) guided by a predictive, hardwareaware energy consumption model. The primary technical contributions include: (1) The E-Model, a calibrated analytical Energy Cost Model that rigorously quantifies layer-wise energy consumption by explicitly incorporating the effects of data movement, precision, and sparsity. (2) The E-Metric (Energy Efficiency Metric), a holistic performance index that quantifies the energyaccuracy Pareto frontier by normalizing efficiency (Tokens/Joule) against model degradation (Perplexity loss). (3) The Adaptive Allocation Algorithm (A-Algorithm), an iterative greedy search approach designed to minimize the E-Metric across the entire network while simultaneously mitigating the cumulative effects of reconstruction error propagation. Implemented within the opensource TinyEdgeLLM Library and validated on a reference edge NPU/CPU platform, EcoLLM achieved a 6.4x compression ratio (from 487MB to 76MB) on the GPT-2 architecture with < 2.0% perplexity degradation, resulting in a 45% reduction in power consumption (from 3.8W to 2.1W) on a reference Raspberry Pi 4 platform.

23 December 2025 in Research Square
Verified
Preprint
Open access
PDF file
Krishna BajpaiKrishna Bajpai
Krishna Bajpai
Modern electronic markets have undergone a structural transition from human-mediated exchange mechanisms to high-frequency adversarial control systems governed by competing algorithmic feedback laws. While traditional economic theory attributes market dislocations to behavioral biases or exogenous shocks, such explanations fail to account for instabilities arising in environments populated by rational, mathematically optimized agents. This paper reframes financial markets as dynamical plants controlled by heterogeneous, interacting controllers operating under latency, information, and impact constraints. We show that instability is not anomalous but emerges endogenously from optimal feedback operating in nonnormal, metastable state spaces. Introducing the concepts of metastable markets, non-normal transient amplification, and a quantitative Control Fragility Index (CFI), we demonstrate how markets can appear statistically stable while remaining critically sensitive to marginal changes in control gain, synchronization, or delay. We further identify a six-part taxonomy of feedbackinduced instabilities—including reflexive resonance, information gradient collapse, and liquidity phase transitions—that arise without manipulation, deception, or irrational behavior. Finally, we argue that market stability must be treated as a firstclass design constraint and propose design-level regulatory interventions grounded in control theory rather than behavioral restriction. Our results suggest that systemic instability is the natural byproduct of local optimality in adversarial feedback environments and that equilibrium existence alone is insufficient to guarantee dynamical safety.

26 December 2025 in Preprints.org
Verified
Preprint
Open access
PDF file
Krishna BajpaiKrishna Bajpai
Krishna Bajpai
The evolution of high-performance computing (HPC) interconnects has produced specialized fabrics such as InfiniBand[1], Intel Omni-Path, and NVIDIA NVLink[2], each optimized for distinct workloads. However, the increasing convergence of HPC, AI/ML, quantum, and neuromorphic computing requires a unified communication substrate capable of supporting diverse requirements including ultra-low latency, high bandwidth, collective operations, and adaptive routing. We present HyperFabric Interconnect (HFI), a novel design that combines the strengths of existing interconnects while addressing their scalability and workload-fragmentation limitations. Our evaluation on simulated clusters demonstrates HFI’s ability to reduce job completion time (JCT) by up to 30%, improve tail latency consistency by 45% under mixed loads and 4× better jitter control in latency-sensitive applications., and sustain efficient scaling across heterogeneous workloads. Beyond simulation, we provide an analytical model and deployment roadmap that highlight HFI’s role as a converged interconnect for the exascale and post-exascale era.

3 March 2026 in SSRN Electronic Journal
Verified
Preprint
Open access
Krishna BajpaiKrishna Bajpai
Krishna Bajpai
VGVedanshi Gupta
Vedanshi Gupta
Deployment of Large Language Models (LLMs) on resource-constrained edge devices, such as IoT systems and mobile AI platforms, is significantly hampered by excessive energy consumption, large memory footprints, and high inference latency. Current methods relying on the sequential application of model compression techniques—namely quantization and pruning—often result in sub-optimal and hardware-unaware configurations. This paper introduces EcoLLM, a novel joint optimization framework designed for achieving ultra-low power LLM inference. EcoLLM systematically determines the optimal layer-wise configurations of mixed-precision bit-widths (bw) and structured sparsity rates (s) guided by a predictive, hardwareaware energy consumption model. The primary technical contributions include: (1) The E-Model, a calibrated analytical Energy Cost Model that rigorously quantifies layer-wise energy consumption by explicitly incorporating the effects of data movement, precision, and sparsity. (2) The E-Metric (Energy Efficiency Metric), a holistic performance index that quantifies the energyaccuracy Pareto frontier by normalizing efficiency (Tokens/Joule) against model degradation (Perplexity loss). (3) The Adaptive Allocation Algorithm (A-Algorithm), an iterative greedy search approach designed to minimize the E-Metric across the entire network while simultaneously mitigating the cumulative effects of reconstruction error propagation. Implemented within the opensource TinyEdgeLLM Library and validated on a reference edge NPU/CPU platform, EcoLLM achieved a 6.4x compression ratio (from 487MB to 76MB) on the GPT-2 architecture with < 2.0% perplexity degradation, resulting in a 45% reduction in power consumption (from 3.8W to 2.1W) on a reference Raspberry Pi 4 platform.1) KeywordsLarge Language Models, Energy Optimization, Mixed-Precision, Quantization, Pruning, Perplexity, Raspberry Pi 4 platform


25 December 2025 in Preprints.org
Verified
Preprint
Open access
PDF file
Krishna BajpaiKrishna Bajpai
Krishna Bajpai
Modern quantitative trading systems implicitly assume that predictive accuracy implies economic control. This paper proves that assumption is false. We formalize a fundamental separation between predictability—an observer-centric statistical property—and controllability—an actuator-centric dynamical property of trading under feedback, frictions, and competition. Using tools from stochastic control, market microstructure, and performative economics, we establish a negative result: high predictive power is neither sufficient for nor monotone in profitability and can, under realistic regimes of convex execution costs and feedback, strictly accelerate P&L erosion. We model markets as closed-loop dynamical systems in which agents’ actions alter the state conditional on which predictions are made. We characterize exploitable alpha via a controllability condition and show that prediction-optimal policies are generically antioptimal under feedback. Finally, we propose a control-theoretic replacement for prediction-centric trading could specify that this replaces MSE-centric optimization to emphasize the practical takeaway.