Keval Parekh

Biggest Learning is to Unlearn - Before Learning Again! Having done multiple Data Mesh implementations in the last 5 years, and having written one academic paper, I kept on looking for opportunities to continue learning in different  dimensions. Very soon I found the right platform to look beyond the corporate boundaries from a different lens and apply  academic frameworks. I embarked on the journey of my 2nd academic paper with other industry experts - Maribel Yasmina Santos, Isabel Ramos , senior researcher in the field of Information Science and Divya Joshi, a seasoned data strategist. The journey  was rewarding as researching and writing academic papers sharpened our critical thinking and analytical skills.  We drew deep insights from the industry trends, emerging technologies, and potential areas for innovation. As we connected with other researchers, professionals and organizations beyond our milieu, our learnings and supportive networks expanded. Our paper along with other high impact submissions in the conference will help us in advancing  thought leadership while providing creative solutions to business challenges in our day jobs. If you want to learn more about what goes behind a successful adoption of Data Mesh across some major implementations, refer our paper - Data Mesh Adoption: A Multi-case and Multi-method Readiness Approach - https://lnkd.in/gfRXRStu This paper won the best paper award at European, Mediterranean, and Middle Eastern Conference on Information Systems. This is a demonstration of results based on  surveys and a set of interviews used in a multi-case and multi-method approach. Researchers and data triangulation were implemented to ensure rigor and arrive at a comprehensive understanding of Data Mesh adoption. Feel free to share your feedback and learnings! Dr MARIA PAPADAKI Marinos Themistocleous Yannis Charalabidis #Datamesh #datamanagement

38

3 Comments

I got to watch this insightful in conversation between Rakesh Ravuri, CTO and Senior VP of Engineering at Publicis Sapient and Vandhna Babu, Principal Analyst at nasscom insights. In this engaging discussion, they delve into the intersection of Agile methodologies and Generative AI, shaping the future of engineering in remarkable ways. Here’s why you should tune in: 1️⃣Agile Evolution with AI: 🎙Discover how Agile methodologies and Generative AI seamlessly intertwine, creating an ecosystem that thrives on efficiency, precision, and innovation. 🎙Gain insights into how these advanced technologies can transform the landscape of Agile, driving evolution in ways previously unimaginable. 2️⃣Unlocking Potential: 🎙Explore how AI tools like Github Copilot and Amazon Code Whisperer are revolutionizing software engineering. 🎙Understand the value of integrating AI into Agile workflows, unlocking greater potential for your teams. Watch the Conversation: https://lnkd.in/gpxdZFR9 Feel free to share this with your network—it’s a must-watch for anyone passionate about Agile, AI, and the future of engineering! 🚀👍 #AgileEngineering #AI #ProductManagement #ThriveAtPS #LifeAtPS

3

A healthy work life balance begins by addressing stress. And addressing stress begins by unlocking the power of your breath. DM me for The Art of Living workshops I am holding that teach tools and techniques for daily use to help de-stress and re-vitalise!!

7

1 Comment

Elevating Banking Customer Service Beyond Prompt-Based Solutions In the dynamic landscape of banking, providing top-notch customer service is paramount. Generative AI (GenAI) has already made significant strides, but recent advancements reveal that Retrieval-Augmented Generation (RAG) based GenAI can take customer service to unprecedented heights. Let's explore how RAG-based GenAI outshines traditional prompt-based GenAI applications. Understanding RAG-Based GenAI RAG-based GenAI integrates retrieval mechanisms with generative models, creating a hybrid system that leverages vast amounts of external information. This combination enhances the AI's ability to generate precise and contextually relevant responses, offering a more refined customer service experience. Advantages of RAG-Based GenAI Accurate and Context-Rich Responses - Precision - Dynamic Updates Enhanced Task Execution - Complex Query Handling - Improved Efficiency Reduced Error Rates - Contextual Understanding - Consistent Performance Addressing Common Challenges While RAG-based GenAI offers remarkable benefits, it also addresses some limitations inherent to traditional GenAI models: Ethical Dilemmas and Value-Based Judgments:  By accessing diverse sources of information, RAG-based systems can provide more balanced perspectives, though human oversight remains crucial. Data Relevance and Predictability:  The retrieval component ensures the model works with relevant, real-time data, enhancing predictability and reliability. Security and Privacy:  Implementing robust security protocols ensures that sensitive customer data is protected during the retrieval and generation processes. Implementation Strategies Custom Solutions:  For banks with unique requirements and vast proprietary data, developing RAG-based GenAI from scratch offers a tailored solution. Model Fine-Tuning:  Leveraging existing RAG models with bank-specific data can quickly enhance performance, balancing cost and efficiency. Explore More Stay tuned for more insights and sample code demonstrating the power of RAG-based GenAI in banking. #GenAI #RAGGenAI #BankingInnovation #CustomerService #AIinBanking

23

One of the most significant challenges India faces is overcoming the computational hurdles in the realm of advanced technologies. To put things into perspective, storing a single model parameter (32-bit float) requires 4 bytes of memory. For a model with 1 billion parameters, this translates to 4 GB of memory. When it comes to training a model—such as a 1 billion parameter Large Language Model (LLM)—the memory requirements escalate dramatically to 80 GB, assuming 32-bit full precision. Here’s the breakdown: model weights require 4 bytes, the Adam optimizer needs 8 bytes, gradients take up 4 bytes, and activation plus temporary memory demands an additional 8 bytes. This totals 20 bytes per parameter, far exceeding the capabilities of consumer-grade hardware. To handle such computational demands, one would need an Nvidia A100 80GB Tensor Core, which costs approximately Rs. 1.4 million, or $0.87 per hour on an A2 virtual machine from a leading cloud service provider. Moreover, the challenge is even more pronounced when dealing with Indic languages. Unlike English, where three characters typically make up a token, Indian languages may have a 1:1 ratio of characters to tokens, further increasing the computational load. The dual challenges of computational power and the high cost of GPUs are substantial for India. While some relief can be achieved through techniques like quantization—reducing integer precision from 32-bit full precision to Bfloat16 to lower memory usage—or through distributed data parallel (DDP) or Fully Sharded Data Parallel (FSDP) frameworks. While classical processors (CPUs/GPUs) are well-suited for general-purpose computing and parallelizable tasks, quantum processors (QPUs) are tailored for highly specialized tasks that leverage quantum mechanics. These two types of processors serve distinct purposes, and as quantum computing technology evolves, it may complement or even surpass classical computing in specific areas. However, classical processors will continue to play an essential role in everyday computing tasks. Quantum technology's potential to break many of the defenses currently securing the internet and other critical applications poses a serious threat to cybersecurity. Most of today’s secure transactions, communications, and online account passwords are protected through encryption algorithms like RSA or SSL/TLS, which rely on the complexity of factoring large numbers into primes—a problem that quantum computers excel at solving. A task that would take a classical computer 100 years could be accomplished in mere seconds with a quantum computer. This not only jeopardizes personal account passwords but also exposes private communications, company data, and even military secrets. As India advances towards becoming a digitally rich nation, it is essential to invest wisely. The National Quantum Mission (NQM) and the initiatives it has set forth are steps in the right direction to address these challenges.

5

Another great course from DeepLearning.AI around use of #langchain to invoke #agents, #modelevaluation and other LLMy things https://lnkd.in/d_sHdyre

60

1 Comment

Insights and merits on creating dynamic APIs Building APIs which are flexible, scalable, and adaptable is fundamental to staying beforehand. 1. Design with flexibility in thoughts: Anticipate converting requirements through designing APIs with modularity and extensibility. Utilize standards like RESTful concepts or GraphQL etc. 2. Versioning strategy: Implement a strong versioning strategy to control adjustments efficiently with out breaking present patron implementations. Whether it is via URI versioning, custom headers, or content negotiation, clear versioning ensures smooth transitions. 3. Parameterization and customization: Empower API consumers with the ability to tailor requests through parameterization and customization. Options like query parameters, headers, or maybe dynamic endpoints allow fine-tuning of responses in keeping with precise necessities. 4. Documentation is king: Comprehensive documentation is non-negotiable. Clear, concise documentation no longer most effective helps ease of integration however also serves as a reference point for builders, reducing friction and accelerating adoption. 5. Error managing and comments: Robust error coping with mechanisms provide treasured feedback to API purchasers, helping in troubleshooting and improving person enjoy. Define clear errors codes, messages, and appropriate status codes to deliver meaningful facts. 6. Security by means of design: Security ought to be ingrained at each layer of the API architecture. Implement enterprise-trendy authentication and authorization mechanisms, at the side of measures like fee limiting and input validation, to protect in opposition to potential threats. 7. Monitoring and analytics: Continuous tracking and analytics offer insights into API performance, usage styles, and ability bottlenecks. Leverage equipment and systems to acquire metrics, detect anomalies, and optimize resource allocation for greater reliability and scalability. By adopting these strategies, one ̉can create APIs that aren't the simplest sturdy and scalable but also agile enough to adapt to changing commercial enterprise necessities and technological improvements. #APIs #SoftwareDevelopment #TechTrends #DeveloperCommunity #Innovation #Technology #TechStack #RestAPIs #GraphQL

10

Do not just use #chatgpt for simple Q&A like give me a sample code, write a query for this etc.. Use it more to explore,gain knowledge on deep topics like architecture, performance & compare b/w various alternatives etc..Keep Q&A with why? If? as if you are sharing your knowledge and clarifying your doubts with a friend or a mentor who is good at that subject.

31

Small language models are gaining popularity in solving specific domain tasks with less cost. Here in this figure we see, with only 1.3 billion parameters Phi-1.5 exhibits a performance on natural language tasks comparable to models 5x larger. Their efficiency, cost and power of customisation provide us motivation in training efficient small language models and deploying SLMs in enterprise infrastructure. One of the key factors behind the success of SLMs is along with the Instruction dataset, we use Direct Preference Optimisation (DPO) on a feedback dataset. Other efficient mechanisms include step-by-step reasoning and recall-then-generate which help to shrink the models and at the same time outperform other certain #LLMs like Llama-7B, Llama2-7B, or Vicuna-13B. However, as research evolves we will continue to answer how small can language models shrink while preserving compelling capabilities through advanced AI training. And that will keep the excitement 😋 going! Reference: https://lnkd.in/dey2-XzX #artificialintelligence #chatgpt #machinelearning #LLM #ai

40

2 Comments

Below are the detailed ETL approaches which can be explored as per requirement. 1. Medallion Architecture, 2. Lambda Architecture, 3. Kappa Architecture, 4. Data Vault Architecture, 5. Kimball’s Dimensional Data Warehouse, 6. Inmon’s Corporate Information Factory, 7. Lakehouse Architecture #databricks #ETL #DataEngineer #Medallion

5

What makes FinTech Product Managers so valuable? 🤔 They drive innovation and bring financial solutions to life by: ✅ Building user-centered strategies ✅ Solving regulatory challenges with agility ✅ Leading tech-focused transformations Imagine becoming the PM companies can’t wait to hire. 🌟 Our bootcamp equips you with the exact skills recruiters look for in top candidates. Here’s what our bootcamp equips you with to make it possible: ✔️ Training in PM methodologies for tech innovation ✔️ Advanced tools to manage compliance and risk ✔️ Career coaching to position you as a FinTech leader 🗓️ First cohort starts: December 12, 2024 This opportunity won’t last long—reserve your spot today! #MDMarketInsights #KaptureX #FinTechPM #ProductManagement #CareerGrowth #FinTechCareers #DigitalInnovation #FutureOfFinance #Blockchain #AIInFinTech #FinTechLeadership #TechTrends #ProfessionalDevelopment #DigitalTransformation #LearnAndGrow

3

AI Enabling BFSI facilitating the process

5

"The Developer's Playbook for Responsible AI in India" – a vital resource for navigating the ethical landscape of AI development. This playbook offers a voluntary framework to help developers identify and mitigate risks associated with building and deploying AI in India. Here are some key highlights from the playbook: ● Comprehensive Risk Mitigation Guides: The playbook offers specific guidance for three types of AI: Discriminative AI Models, Generative AI Models, and AI Applications. ● Focus on the AI Lifecycle: Each guide breaks down potential risks and mitigation strategies across all stages of development, from conception and data collection to deployment, monitoring, and maintenance. ● Practical and Actionable Advice: The playbook goes beyond theoretical principles by providing clear, step-by-step prompts for developers to implement responsible AI practices. ● Emphasis on Data Privacy: Recognizing the importance of data protection, the playbook extensively covers compliance with the Digital Personal Data Protection Act, 2023. ● Transparency and Explainability: The playbook stresses the need for clear documentation, explainable outputs, and stakeholder engagement to foster trust in AI systems. Why is this important? As AI rapidly transforms various sectors in India, ensuring responsible development is crucial. This playbook empowers developers to build AI solutions that are not only innovative but also ethical, safe, and beneficial for all. Let's work together to build a thriving and responsible AI ecosystem in India! #ResponsibleAI #AIinIndia #nasscom #AIethics #AIDevelopment #DataPrivacy

18

2 Comments

One common mistake we often make in life is retreating to our comfort zone too quickly. While celebrating success is crucial, staying on the path of growth is equally essential. After 4 weeks of gap, I tried to do some learning activities today and realized how distant it felt. 🤔 Hopeful to not repeat this and strive to learn every day , maybe just for 5 minutes .. 🤞 Be Continuous... Be Confident..... Be Prepared.... 📖 📚 #ContinuousLearning #ComfortZoneIsDangerous #LearnEveryDay

16

1 Comment

🔴 Moving data from Fusion SaaS layer to on-prem / cloud staging is a common requirement across organizations. 🔶 There are various approaches available for this outbound data movement from SaaS layer. Native capabilities like BIP, BICC, REST end points etc. can be leveraged for this. ( more details : https://lnkd.in/dTkfgy9c). 🔶 Building a staging layer within the cloud using Autonomous Database is a good approach as it can orchestrate the movement of data to third parties ( bulk, or via APIs) and on prem downstream systems in a secured way. 🔶 A high level approach to solve this use case can be as follows: 1️⃣ Set up BICC to connect to OCI Object Storage ( refer : https://lnkd.in/dTejbt4b) 2️⃣ Using Data Load in Autonomous database, load the data from Object storage. If any transformation to be applied within the pipeline, leverage ODI Web which is embedded within Autonomous database ✴ Use Cases for Moving Fusion Data: 1️⃣ Reporting & Analytics: Create dedicated data warehouses or data marts for advanced analytics and reporting, tailored to specific business needs. 2️⃣ Data Science & ML: Extract relevant Fusion ( CX/ ERP) data to train machine learning models for predictive analytics, forecasting, or anomaly detection. 3️⃣ Custom Applications: Build custom applications that leverage Fusion data, integrated seamlessly with the dedicated database. 4️⃣ Archiving & Compliance: Maintain historical Fusion data in a separate database for compliance or long-term storage purposes. 5️⃣ Data Consolidation: Integrate data from multiple Fusion instances or other sources into a centralized database for unified analysis. #datamovement #Fusion #ADW #orchestration #quickwins

32