VVVinay Venkatesh

Vinay Venkatesh

*This is an algorithmically generated author record based on publicly available information.

Vinay Venkatesh has a total of 20 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Shape Prior Fusion for Efficient 3D Reconstruction

    by Vamsidhar R. Kamanuru, Vinay Venkatesh, Hrishikesh Tawade, Sai Charan Dekkata, Vincent Koc

    Creating accurate 3D models from images is a challenging problem in computer vision, especially for applications like autonomous driving, virtual reality, and mobile computing. Traditional Multi-View Stereo (MVS) methods rely on geometry to reconstruct shapes but often struggle with missing details due to occlusions or textureless surfaces. In this work, we introduce a learning-based approach that enhances MVS by integrating deep priors, allowing for more complete and accurate 3D reconstructions with fewer input images. Our method refines fine details using masked silhouette and depth map losses, ensuring better shape recovery. Tested on the ShapeNet dataset, our model achieves a Chamfer Distance of 0.018 and an IoU of 0.80, outperforming classical MVS while also improving surface coverage to 91.1% and density score to 85.7%. By combining deep learning with traditional geometry, our approach makes 3D reconstruction more reliable and efficient, opening new possibilities for real-world applications.

  • Adaptive Hybrid Agents for CRM: Combining Rule Engines, LLMs, and Feedback Loops for Reliable Automation

    by Nikita Kothari, Vinay Venkatesh, Varun Joshi, Lav Kumar

    As enterprise systems evolve from deterministic automation toward adaptive agentic intelligence, the challenge of balancing creativity, compliance, and reliability has become central to Customer Relationship Management (CRM) automation. This research presents the Adaptive Hybrid Agent (AHA) — a unified architecture that combines symbolic rule reasoning, large language model (LLM) inference, and continuous feedback-driven adaptation. The AHA architecture features a meta-controller that dynamically orchestrates between symbolic, neural, and hybrid reasoning paths according to context complexity, confidence thresholds, and feedback gradients derived from user interaction and system metrics. By employing multi-modal feedback including rule validation outcomes, engagement signals, and hallucination scores, the framework achieves self-calibrating behavior that improves both accuracy and adaptability under domain drift. Empirical evaluation across CRM workflows such as campaign generation, text-to-SQL analytic reasoning, and policy-aware message composition demonstrates significant improvements in factual grounding (+57%), task success rate (+14%), and engagement uplift (+25%) relative to static or single-modality baselines. These findings suggest that adaptive hybrid architectures enable reliable, auditable, and continually self-improving CRM agents, capable of aligning generative reasoning with enterprise governance.

  • LLM-Augmented Semantic Resonance System (SRR): A Unified Triple-Objective Architecture for Deep Sequential Recommendation

    by Vinay Venkatesh, Devdas Gupta, Lav Kumar, Ganesh Harke

    Sequential recommenders have recently achieved strong gains by modeling the temporal structure of user behavior, yet most systems still rely on shallow item representations that capture co-occurrence statistics rather than the underlying semantics of content. This paper proposes the Semantic Resonance System (SRR), a unified architecture that integrates deep semantic signals from a large language model (LLM) with standard behavioral embeddings for sequential recommendation. SRR uses a domain-adapted Llama 2 13B model, fine-tuned via ParameterEfficient Fine-Tuning (LoRA), to produce high-fidelity semantic item embeddings that are fused with behavioral embeddings through a Dynamic Cross-Attention module. A central component of SRR is the Semantic Resonance Optimization (SRO) loss, which jointly optimizes (i) a Dense All-Action (DAA) loss for high-recall retrieval, (ii) a Binary Cross-Entropy (BCE) loss for calibrated scores, and (iii) a Contrastive Alignment loss $\mathcal{L}_{C A}$ that explicitly aligns semantic and behavioral latent spaces. On the large-scale MovieLens $\mathbf{2 5 M}$ dataset, SRR achieves a $\mathbf{1 6. 4 \%}$ improvement in NDCG@10 over a strong transfer-learning baseline, while also improving diversity, catalog coverage, and cold-start performance. A latency-effectiveness study further shows that SRR can be configured along a smooth Pareto frontier between NDCG@10 and inference cost, supporting deployment under different resource budgets.

  • DP-Prox: A Robust and Differentially Private Framework for Federated Instruction Tuning of Small LLMs on 8 GB Edge Devices

    by Vinay Venkatesh, Yichen Liu

    This paper introduces DP-Prox, an end-to-end framework for robust and differentially private federated instruction tuning of small Large Language Models (LLMs) on 8 GB edge devices. DP-Prox is instantiated using the 3.8 billionparameter Phi-3-mini model with 4-bit Quantized Low-Rank Adaptation (QLoRA), and combines (i) FedProx-style proximal regularization on low-rank adapters to stabilize optimization under highly Non-IID smart-home data, and (ii) Differentially Private SGD (DP-SGD) applied directly to adapter gradients to provide user-level $(\epsilon, \delta)$-differential privacy against an honest-but-curious server. In a realistic IoT instruction-following benchmark with Dirichlet-skewed clients ($\alpha=0.3$), DP-Prox improves F1-score from 0.80 (Edge-FIT with FedAvg) to 0.82 while reducing convergence rounds from $T=50$ to $T \approx 40$. Under a strict privacy budget $\left(\epsilon=5, \delta=10^{-5}\right)$, DP-Prox maintains an F1-score of 0.77 with only $+15 \%-20 \%$ local overhead, and preserves the communication efficiency of QLoRA-based federated training. Qualitative analyses further show that DP-Prox yields more consistent instruction-following behavior across heterogeneous homes, especially on multi-device routines and privacy-sensitive queries. Together, these results position DP-Prox as a practical and regulation-ready blueprint for deploying small, edge-resident LLMs in smart-home IoT environments.


  • Open Access

    DP-Prox: A Robust and Differentially Private Framework for Federated Instruction Tuning of Small LLMs on 8 GB Edge Devices

    by Vinay Venkatesh, Yichen Liu

    This paper introduces DP-Prox, an end-to-end framework for robust and differentially private federated instruction tuning of small Large Language Models (LLMs) on 8 GB edge devices. DP-Prox is instantiated using the 3.8 billionparameter Phi-3-mini model with 4-bit Quantized Low-Rank Adaptation (QLoRA), and combines (i) FedProx-style proximal regularization on low-rank adapters to stabilize optimization under highly Non-IID smart-home data, and (ii) Differentially Private SGD (DP-SGD) applied directly to adapter gradients to provide user-level (ϵ, δ)-differential privacy against an honestbut-curious server. In a realistic IoT instruction-following benchmark with Dirichlet-skewed clients (α = 0.3), DP-Prox improves F1-score from 0.80 (Edge-FIT with FedAvg) to 0.82 while reducing convergence rounds from T = 50 to T ≈ 40. Under a strict privacy budget (ϵ = 5, δ = 10 −5), DP-Prox maintains an F1-score of 0.77 with only +15%-20% local overhead, and preserves the communication efficiency of QLoRA-based federated training. Qualitative analyses further show that DP-Prox yields more consistent instruction-following behavior across heterogeneous homes, especially on multi-device routines and privacy-sensitive queries. Together, these results position DP-Prox as a practical and regulation-ready blueprint for deploying small, edge-resident LLMs in smart-home IoT environments.

  • Domain Focused Sequence Model Architecture for Predicting Long Term User Interests

    by Vinay Venkatesh, Akshay Mittal, Nikita Kothari, Varun Joshi

    This paper introduces a novel approach to sequential user modeling by developing distilled sequence model architectures for specific user clusters. While generic, interest-based models have become a cornerstone of personalized recommendation systems, they often fail to provide high-quality recommendations for “power users” with highly specific interests, as their recommendations can be diluted by broader user behavior. To address this, we present a framework for creating specialized transformer-based sequence models tailored to individual domains, such as horror movies, music, and sports. Our method, built upon the principles of the Pinner-Former architecture, leverages domain-focused training datasets to generate distilled user embeddings. This approach offers several advantages, including enabling highly focused retrieval and requiring less computational resources, which results in a smaller, more efficient model that's easier to deploy in a production environment. However, this method also presents challenges, such as the difficulty of acquiring high-quality domain-specific data and a lack of scalability for a large number of verticals, as a separate model is needed for each domain. We demonstrate a 19.9% improvement in Recall@10 with 70-75% computational efficiency improvements, showing that this specialized approach significantly improves personalization for power users while enabling more efficient deployment.


  • Adaptive Weighted Loss for Sequential Recommendations on Sparse Domains

    by Akshay Mittal, Vinay Venkatesh, Krishna Kandi, Shalini Sudarsan

    The effectiveness of single-model sequential recommendation architectures, while scalable, is often limited when catering to “power users” in sparse or niche domains. Our previous research, PinnerFormerLite, addressed this by using a fixed weighted loss to prioritize specific domains. However, this approach can be sub-optimal, as a single, uniform weight may not be sufficient for domains with very few interactions, where the training signal is easily diluted by the vast, generic dataset. This paper proposes a novel, data-driven approach: a Dynamic Weighted Loss function with comprehensive theoretical foundations and extensive empirical validation. We introduce an adaptive algorithm that adjusts the loss weight for each domain based on its sparsity in the training data, assigning a higher weight to sparser domains and a lower weight to denser ones. This ensures that even rare user interests contribute a meaningful gradient signal, preventing them from being overshadowed. We provide rigorous theoretical analysis including convergence proofs, complexity analysis, and bounds analysis to establish the stability and efficiency of our approach. Our comprehensive empirical validation across four diverse datasets (MovieLens, Amazon Electronics, Yelp Business, LastFM Music) with state-of-the-art baselines (SIGMA, CALRec, SparseEnNet) demonstrates that this dynamic weighting system significantly outperforms all comparison methods, particularly for sparse domains, achieving substantial lifts in key metrics like Recall@10 and NDCG@10 while maintaining performance on denser domains and introducing minimal computational overhead.


  • Open Access

    Transfer Learning for Efficient Domain Adaptation in Sequential Recommendation Systems

    by Vinay Venkatesh, Akshay Mittal, Sankaranainar Parmsivan, Kathiresan Jayabalan

    The effectiveness of sequential modeling in personalized recommendation systems is often hindered by the challenge of providing high-precision recommendations to "power users" with concentrated, niche interests. Generic, single-model architectures tend to dilute these specific signals within a vast sea of general user behavior. While our previous research introduced domain-specific models and weighted loss functions as solutions, these approaches either lack scalability or may still struggle with sparse domains. This paper proposes a novel and more efficient framework based on transfer learning to combine the benefits of both. We introduce a methodology that fine-tunes a single, pre-trained PinnerFormerbased foundation model on smaller, domain-specific datasets. This approach bypasses the need to train a new, large model from scratch for each domain, significantly reducing computational costs while achieving superior performance. Our empirical validation on the MovieLens 25M Dataset shows that our fine-tuned model achieves substantial lifts in key metrics like Recall@10 and NDCG@10 for both "Horror" and "Action" movie domains, demonstrating its generalizability and efficiency.

  • Open Access

    Domain Focused Sequence Model Architecture For Predicting Long Term User Interests

    by Vinay Venkatesh, Akshay Mittal, Nikita Kothari, Varun Joshi

    This paper introduces a novel approach to sequential user modeling by developing distilled sequence model architectures for specific user clusters. While generic, interestbased models have become a cornerstone of personalized recommendation systems, they often fail to provide high-quality recommendations for "power users" with highly specific interests, as their recommendations can be diluted by broader user behavior. To address this, we present a framework for creating specialized transformer-based sequence models tailored to individual domains, such as horror movies, music, and sports. Our method, built upon the principles of the PinnerFormer architecture [1], leverages domain-focused training datasets to generate distilled user embeddings. This approach offers several advantages, including enabling highly focused retrieval and requiring less computational resources, which results in a smaller, more efficient model that's easier to deploy in a production environment. However, this method also presents challenges, such as the difficulty of acquiring high-quality domain-specific data and a lack of scalability for a large number of verticals, as a separate model is needed for each domain. We demonstrate that, despite these tradeoffs, this specialized approach can significantly improve personalization for power users, leading to more relevant recommendations and a better user experience.

  • Open Access

    Bridging Context, Statistics, and Practice: A Multi-Dimensional Framework for Responsible LLM Evaluation and Selection

    by Nitin Kumar, Shiva Kumar Ramavath, Vinay Venkatesh

    Current AI model evaluation approaches employ uniform weighting schemes that fail to capture context-specific requirements and stakeholder priorities, leading to suboptimal model selection decisions in real-world deployments. Traditional methods prioritize raw performance metrics while neglecting critical dimensions such as fairness, privacy, and robustness, resulting in misaligned model choices for safetycritical, creative, and social applications. We introduce ARIA (AI Responsibility and Impact Assessment), a novel contextadaptive evaluation framework that dynamically weights five responsibility dimensions-performance, fairness, robustness, privacy, and sustainability-based on application context and stakeholder profiles. ARIA incorporates uncertainty quantification, cascading penalty mechanisms for critical weaknesses, and multi-objective optimization to provide comprehensive model assessment beyond traditional accuracyfocused approaches. Our empirical evaluation across three state-of-the-art language models (LLaMA 3.3 70B, Claude 3.5 Sonnet, GPT-4o) in critical systems, creative tools, and social platform contexts demonstrates ARIA's superiority over conventional methods. Results show low correlations (Kendall's τ =-0.67 to 0.67) between ARIA and traditional rankings, indicating that uniform approaches misrepresent model suitability in up to 60% of scenarios. Statistical analysis confirms significant performance differences with 95% confidence intervals and robust rank stability (>0.89) underweight perturbations. ARIA provides actionable decision support through contextstakeholder matrices, enabling organizations to align model selection with specific requirements and risk tolerance. The framework addresses critical gaps in responsible AI deployment, offering practitioners a scientifically rigorous methodology for context-appropriate model selection that balances performance excellence with multi-dimensional responsibility considerations, supporting more informed and ethical AI system deployment decisions.