Capgemini

• Designed, Developed, Deployed, Optimized and built global tools and products for the Financial,  Hospitality, Aerospace and Automotive domains.

Associate Consultant

Linkon.AI

• Provided end to end solutions for Software/Web/AI with real-time Deployment. Built products according to business needs.

• Software Development and Web Development [Front End/Back End/Tech Stack (MERN)] to produce high-quality website development solutions.

• Advanced Machine Learning to build prediction model dashboards and solutions for various problems.

• Advanced Deep Learning, Transformers, Artificial Neural Networks, Convolutional Neural Networks, Recurrent Neural Networks, Self Organizing Maps, Generative Adversarial Networks, and autoencoders.

• Advanced Computer Vision to building dashboards for Object Detection/Tracking/Recognition, Image Segmentation (Instance/Semantic), Image/Video classification, Sentiment Recognition, Tracking and Motion Analysis, Neural Style Transfer, Optical Character Recognition and Autonomous Vehicles to solve various problems.

• Advanced Natural Language Processing, Transformers, Sentiment Analysis, Semantic Similarity in Text, Named Entity Recognition, Text Summarization, Subjective Sentiment Analysis, Topic Modelling, Information Extraction and Analysis, building Chatbots using Indian languages.

• Launching a New Company, Business Presentations, Venture Capital (How to raise money), Security/Legal Structures to protect you, Investor Valuations Topics & Dilution, Importance of Contracts, Financial Analysis, Managing Cash Flow, Financial Capital and Securities Law, Projecting Financial Statements, Due Diligence & Data Sources, Modeling & Valuation, Initial Public Offering (IPO) & Valuation, Management Analytical Frameworks, Alternatives and Turnarounds, Crucial Leftover Content/ Goal setting workshops, Closing remarks. 

• Product Management and Development [Internal, B2B, B2C] Including Ideas and User needs, Competitive and Market Analysis, Customer Development, Designing and Running Experiments, Conceptualizing the Solution, Defining Success and Measuring Results, Project Management, Worked with People and Stakeholders.

Founder

worxogo

• Applied Software Engineering, Data Science and Engineering, Data Analytics, Behavior Science and Design, Machine Learning, Artificial Intelligence to solve various real-world business problems using the latest tools and technologies.

Software Engineer (Machine Learning/AI)

Experienced Machine Learning Engineer, Data Scientist and Engineering professional with a demonstrated history of working in Software Engineering background with specialized in Product Management and Development with a good understanding of analytics have the ability to understand business problems and solve it using the latest tools and technologies. Skilled in Deep Learning, Natural Language Processing, Computer Vision, Generative AI, Autonomous Vehicles, AR/VR and Blockchain!

Technical Skills:

•  Programming - Python, C/C++, R, SQL, MATLAB, LaTeX, JavaScript
•  Operating System - Linux [Ubuntu/Fedora/Kali], Mac OS, Windows
•  Web Development - Front End [HTML/CSS/JavaScript], Back End [Python/PHP/SQL], Tech Stack [MERN]
•  Data Engineering - [Ingestion, Collection, Preparation (ETL)], Big Data, Hadoop, Python/PySpark, Spark with AWS & Azure, Spark SQL]
•  Data Analytics - [Data Preparation, Visualization & Preprocessing [Python/R/SQL/Tableau/Excel], API Development
•  Data Science - Python/PySpark, R, SQL, Machine Learning & Artificial Intelligence (Model development and deployment)
•  Machine Learning / Deep Learning (NLP/CV/AV) Frameworks - Scikit Learn, PyTorch, Theano, TensorFlow, Keras, Caffe, OpenCV
•  Database & Cloud Platform - MySQL, PostgreSQL, Snowflake [Amazon Web Services, Google Cloud Platform, Microsoft Azure, Palantir]
•  Tools & Software - PyCharm, Visual Studio Code, Sublime Text, MySQL, JIRA, Docker, Kubernetes, RStudio, Jupyter, Palantir
•  Certifications - 400+ with focusing on Software Development and Artificial Intelligence
•  Others - Business Intelligence, Version Control (Git), Jira, API Development, System Design and Algorithm Development, Parallel Computing, Generative AI

Management Skills:

•  Problem Solving & Decision Making
•  Project and Team Management
•  Strategic Planning & New Business Development

I'm currently exploring new career opportunities and can be reached either through this profile or email at: gyanachandpanigrahi1@gmail.com

llm.c: 1,000 lines of pure C code that implements a GPT-2 and training pipeline from scratch. No python, no frameworks, no dependencies. Just math talking to bare metal. Karpathy makes AI education not only accessible, but also elegant. I'm reading through the code like a work of art. 

https://lnkd.in/g-uT-aPx

Gyanachand Panigrahi

Data preprocessing is critical for building effective RAG systems. Our new short course, Preprocessing Unstructured Data for LLM Applications, taught by Matt Robinson of unstructured.io, demonstrates important but sometimes overlooked aspects of RAG systems:

- How to extract and normalize content from diverse formats like PDF, Powerpoint, and HTML to expand your LLM's knowledge
- Enriching data with metadata to enable more powerful retrieval and reasoning
- Applying document layout analysis and vision transforms to process embedded images and tables

Then you’ll apply all these skills and build a RAG bot that draws from a corpus that includes PDF, PowerPoint, and Markdown documents.

Please sign up here: https://lnkd.in/g84D4dJg

Top 9 Architectural Patterns for Data and Communication Flow 
. 
. 
🔹 Peer-to-Peer 
The Peer-to-Peer pattern involves direct communication between two components without the need for a central coordinator. 
 
🔹 API Gateway 
An API Gateway acts as a single entry point for all client requests to the backend services of an application. 
 
🔹 Pub-Sub 
The Pub-Sub pattern decouples the producers of messages (publishers) from the consumers of messages (subscribers) through a message broker. 
 
🔹 Request-Response 
This is one of the most fundamental integration patterns, where a client sends a request to a server and waits for a response. 
 
🔹 Event Sourcing 
Event Sourcing involves storing the state changes of an application as a sequence of events. 
 
🔹 ETL 
ETL is a data integration pattern used to gather data from multiple sources, transform it into a structured format, and load it into a destination database. 
 
🔹 Batching 
Batching involves accumulating data over a period or until a certain threshold is met before processing it as a single group. 
 
🔹 Streaming Processing 
Streaming Processing allows for the continuous ingestion, processing, and analysis of data streams in real-time. 
 
🔹 Orchestration 
Orchestration involves a central coordinator (an orchestrator) managing the interactions between distributed components or services to achieve a workflow or business process. 
 
-- 
Subscribe to our weekly newsletter to get a Free System Design PDF (158 pages): https://bit.ly/3KCnWXq 
 
#systemdesign #coding #interviewtips 
.

How Netflix Really Uses Java?

Netflix is predominantly a Java shop.

Every backend application (including internal apps, streaming, and movie production apps) at Netflix is a Java application. 

However, the Java stack is not static and has gone through multiple iterations over the years.

Here are the details of those iterations:

1 - API Gateway
Netflix follows a microservices architecture. Every piece of functionality and data is owned by a microservice built using Java (initially version 8)

This means that rendering one screen (such as the List of List of Movies or LOLOMO) involved fetching data from 10s of microservices. But making all these calls from the client created a performance problem.

Netflix initially used the API Gateway pattern using Zuul to handle the orchestration.

2 - BFFs with Groovy & RxJava
Using a single gateway for multiple clients was a problem for Netflix because each client (such as TV, mobile apps, or web browser) had subtle differences.

To handle this, Netflix used the Backend-for-Frontend (BFF) pattern. Zuul was moved to the role of a proxy

In this pattern, every frontend or UI gets its own mini backend that performs the request fanout and orchestration for multiple services.

The BFFs were built using Groovy scripts and the service fanout was done using RxJava for thread management.

3 - GraphQL Federation
The Groovy and RxJava approach required more work from the UI developers in creating the Groovy scripts. Also, reactive programming is generally hard.

Recently, Netflix moved to GraphQL Federation. With GraphQL, a client can specify exactly what set of fields it needs, thereby solving the problem of overfetching and underfetching with REST APIs.

The GraphQL Federation takes care of calling the necessary microservices to fetch the data. 

These microservices are called Domain Graph Service (DGS) and are built using Java 17, Spring Boot 3, and Spring Boot Netflix OSS packages. The move from Java 8 to Java 17 resulted in 20% CPU gains.

More recently, Netflix has started to migrate to Java 21 to take advantage of features like virtual threads.

--
Subscribe to our weekly newsletter to get a Free System Design PDF (158 pages): https://bit.ly/3KCnWXq

#systemdesign #coding #interviewtips
.


OpenAI just released Majorly improved GPT-4 Turbo with Vision🚀

Devin, built by cognition_labs, is an AI software engineering assistant powered by GPT-4 Turbo that uses vision for a variety of coding tasks.

Make Real, built by tldraw, lets users draw UI on a whiteboard and uses GPT-4 Turbo with Vision to generate a working website powered by real code.

Vision requests can now also use JSON mode and function calling.
https://lnkd.in/gAPP_W4Y

Also, two awesome webinars to get more hands-on and add projects to your resume: 
"AI-Powered Personalized Shopping with OpenAI" https://bit.ly/4ao4V5Q
"Build Vector Embeddings for Video via OpenAI CLIP": https://bit.ly/3PT3y6Y

#innovation #chatgpt #artificialintelligence #technology #machinelearning 

Our next-generation AI model Gemini 1.5 Pro is now available in public preview on Google Cloud's #VertexAI platform. 🤝

Its long-context window is already helping businesses analyze large amounts of data, build AI-powered customer service agents and more. 

It also supports processing audio inputs - including music, speech, and even the audio portion of videos. This means it can give high-quality transcriptions or be used to search and analyze multimodal content.

Developers can access Gemini 1.5 Pro via the Gemini API in public preview - now available in 180+ countries.

Find out more. → https://dpmd.ai/43QZrOt #GoogleCloudNext

Introducing the next generation of the Meta Training and Inference Accelerator (MTIA), the next in our family of custom-made silicon, designed for Meta’s AI workloads.

Full details ➡️ https://go.fb.me/pnyh2l

MTIA is a part of our growing investment in AI infrastructure to provide the most efficient architecture for Meta’s unique AI workloads and improve our ability to provide the best experiences for our users around the world.

Apple Pay Vs Google Pay

The diagram below shows the differences. Both approaches are very secure, but the implementations are different. To understand the difference, we break down the process into two flows.

1 Registering your credit card flow
2 Basic payment flow

1. The registration flow is represented by steps 1~3 for both cases. The difference is:
Apple Pay: Apple doesn’t store any card info. It passes the card info to the bank. Bank returns a token called DAN (device account number) to the iPhone. iPhone then stores DAN into a special hardware chip.
Google Pay: When you register the credit card with Google Pay, the card info is stored in the Google server. Google returns a payment token to the phone.

2. When you click the “Pay” button on your phone, the basic payment flow starts. Here are the differences:
Apple Pay: For iPhone, the e-commerce server passes the DAN to the bank.
Google Pay: In the Google Pay case, the e-commerce server passes the payment token to the Google server. Google server looks up the credit card info and passes it to the bank.

In the diagram, the red arrow means the credit card info is available on the public network, although it is encrypted. 

Over to you: Apple needs to discuss the DAN details with banks. It takes time and effort, but the benefit is that the credit card info is on the public network only once. If you are an architect and have to choose between security and cost, which solution do you prefer?

--
Subscribe to our weekly newsletter to get a Free System Design PDF (158 pages): https://bit.ly/3KCnWXq

#systemdesign #coding #interviewtips
.

New open model from Mistral AI! 🧠 Yesterday night, Mistral released Mixtral 8x22B a 141B MoE via magnet link. 🔗🤯

What we know so far:
🧮 141B MoE with ~40B active
📜 context length of 65k tokens.
🪨 Base model can be fine-tuned
👀 ~260GB VRAM in fp16, 73GB in int4
📜 Apache 2.0, according to their discord
🤗 Available on Hugging Face (community)
🔠 Similar tokenizer as previous models
❌ no official evals or performance numbers
❌ no information about pertaining or dataset
❌ no information about Language support

Model: https://lnkd.in/eyCqeSB8

Big Kudos to Mistral! Let's see what the evaluations will tell. 🚀

🚀 𝐂𝐡𝐞𝐚𝐭𝐬𝐡𝐞𝐞𝐭: 𝐆𝐞𝐧𝐞𝐫𝐚𝐭𝐢𝐯𝐞 𝐀𝐈 𝐔𝐬𝐞 𝐂𝐚𝐬𝐞𝐬 + 𝐔𝐬𝐞𝐟𝐮𝐥 𝐓𝐨𝐨𝐥𝐬

Save this useful AI Tools to boost your work:

ChatGPT
Perplexity
Claude AI
Wand
Notion
Tableau
Adobe Firefly
Midjourney
Pika

Harness the power of AI and elevate your business performance!

#ai #chatgpt #cheatsheet #generativeai

Gemma can now code!🤯 🔔 Google DeepMind just released Code Gemma, a collection of specialized open code models. Code Gemma comes in two different sizes 2B & 7B, excellent for on-device code completion. 🧑🏻‍💻

𝗧𝗟;𝗗𝗥;
🧮 2B & 7B with 8192 context
🛫 initialized from Gemma Base
🔠 Trained on 500B additional Tokens (web, code & math)
🛠 Fine-tuned Instruct with SFT & RLHF
🐍 2B achieves 27% on HumanEval and 7B-it 52%
✅ Commercial use allowed
📱 Optimized for on-device code completion
🤗 Available on Hugging Face

Blog: https://lnkd.in/esEUu_rW
Models: https://lnkd.in/e8HGrz35
Report: https://goo.gle/codegemma
Google Blog: https://lnkd.in/eZhci47b

huggingface.co • 5 min read

CodeGemma - an official Google release for code LLMs

Our generative technology Imagen 2 can now create short, 4-second live images from a single prompt. 🖼

Here’s what you need to know:
🔵 It’s available to use in Google Cloud's #VertexAI platform to help enterprise workloads
🔵 We’re adding new advanced photo editing features, including inpainting and outpainting, to make it easier to remove unwanted elements, include new elements, and expand the borders of images to create a wider field of view
🔵 It comes with SynthID, a cutting-edge tool for watermarking and identifying AI-generated content

Find out more. → https://dpmd.ai/43QZrOt #GoogleCloudNext

I am really excited to introduce our first custom-designed CPU, Google Axion Processors. Our ability to design for the most-demanding Cloud workloads has allowed us to build a an Arm CPU that is:
- 30% better performance than the fastest general-purpose Arm-based instances available in the cloud today, and
- up to 50% better performance and up to 60% better energy-efficiency than comparable current-generation x86-based instances.

Our Axion CPUs builds on a long tradition of Google custom hardware, including TPUs for ML workloads, VCUs for video transcoding, and three generations of Tensor chips for mobile devices.

cloud.google.com • 6 min read

Introducing Google’s new Arm-based CPU | Google Cloud Blog

What if you care about finetuning LLMs for classification? Are encoder-style transformers like BERT and RoBERTa, etc, still the way to go? The recent "Label Supervised LLaMA Finetuning" paper shows that you can get really good performance by finetuning decoder-style models such as 7B or 13B Llama 2 on classification tasks. And they are even better if you remove the causal attention mask: "Label Supervised LLaMA Finetuning"

Caveat: OK, but it's important to note that a 7B Llama 2 model is also 70x larger than BERT and thus significantly more expensive to run. Therefore, it's a decision to consider if you prioritize accuracy over computational efficiency.

PS: I shared similar experiments a few years ago. I found that finetuning a 560 million parameter BLOOM model in the same fashion as they described and got 87% accuracy on the IMDB dataset.

However, this was far from the 96% classification accuracy you could get with Roberta. For instance, researchers who took my DL course got up to 96% accuracy with a 358 M parameter RoBERTa model.

So, the obvious takeaway of this paper is that it's just a matter of the model size, surprise, surprise.

A Good List of Tech Articles to Read this Week:

1) How to Answer a System Design Interview Problem: https://lnkd.in/gkhnZU77
Author: Ashish Pratap Singh

2) How Amazon S3 Achieves 99.999999999% Durability: https://lnkd.in/g3s9Nif7
Author: Neo K.

3) Build your own Chess Game: https://lnkd.in/gp8fcW8F
Author: John Crickett

4) How to Self-Manage Even if You Have a Manager: https://lnkd.in/gBM4Gq9F
Author: Irina Stanescu

5) 3 software development principles I wish I knew earlier in my career: https://lnkd.in/gWfNVmAN
Author: Petar Ivanov

6) How LinkedIn Uses Caching to Serve 5M Profile Reads/Sec: https://lnkd.in/gwnJdXvT
Author: Saurabh Dashora

7) How to Use Your Mentor Effectively: https://lnkd.in/ghngmhkm
Author: Raviraj Achar

Also, checkout Alexandre Zajac's "Hungry Minds" for curated list of tech articles every week: https://lnkd.in/gzYPd63r

Command R+ just landed in HuggingChat 💬

This is a historic moment as Command R+ is the first open-weights model to beat GPT4 in the Chatbot Arena. 🤯

Kudos Cohere you are advancing AI for all of humanity.

Try it out now, running on fast optimized inference on Hugging Face infra:

hf.co/chat 

The moat of software AI agents is not the thin wrapper layer (Devin, SWE-Agent), but the underlying LLM. Instead of benchmarking the wrapper, I think SWE-Bench is excellent for evaluating coding LLMs instead:

Hold the agent layer fixed and vary only the LLM backend. Provide all tools: LLM-friendly command line, editor, web browser. Specify a maximum token budget and internet query budget. 

How do GPTs, Gemini's, Claude's, Command R, Mistral, Llama compare with each other on solving real-world software engineering problems, with human-level tool access? This is a MUCH more realistic benchmark than HumanEval. It tests an LLM's ability to search StackOverflow, read documentations, self-reflect, self-correct, and carry out long-term consistent plans.

In the Voyager (LLM playing Minecraft) paper, we did such an experiment where we held the Voyager algorithm fixed and swapped out the GPT-4 backend (orange curve) with 3.5 (green curve). It makes a world of difference.

With a bad LLM, neither Devin nor Voyager matters - you won't get good performance any way. Conversely, with a much stronger LLM (eyeing GPT-5), we can probably get to L4-L5 software automation with just a bit of extra effort. 

https://lnkd.in/gJpp6tzC
LinkedIn post on Voyager: https://lnkd.in/gBB2AtHa

Copilot for Microsoft 365 Success Kit ✨

Download the complete set of assets, now ready for delivery to your team 🗣

These resources help you accelerate onboarding and adoption with your teams beginning their skilling journey with Copilot for Microsoft 365

Copilot for Microsoft 365 Success Kit includes:
✅ Start guide
✅ Implementation summary guide for leaders
✅ User enablement guide
✅ Technical readiness
✅ Scenario library
✅ Accelerate copilot enablement with Viva
✅ Training content map (This image)
✅ Creating an AI Council Guide
✅ Stakeholder engagement map
✅ Onboard and engagement kit
✅ Copilot icon download

🔗 Website: https://lnkd.in/gUxbyKKe

🔗 Download the full kit here (Direct download): https://lnkd.in/gUxbyKKe

#microsoft #copilot #microsoftcopilot #techcnology #ai #artificiallintelligence #microsoft365 #onboarding #adoption


Today at Google Cloud Next, we announced a number of significant enhancements to our Vertex AI platform. A subset of these include:
- Gemini 1.5 Pro with the world's largest context window.
- Imagen 2.0 image generation with 4-second live images
- The ability to ground results with Google Search, substantially improving model accuracy.
- A number of features to ease the path to production for models with side by side evaluation and prompt management services.

Please see the blog post for the details.

Taplio

Gyanachand Panigrahi's Linkedin Analytics

Gyanachand Panigrahi

What is Gyanachand talking about?

Gyanachand Panigrahi's Best Posts (last 30 days)

Want to drive more opportunities from LinkedIn?

Taplio

.css-1jgreyo{background-image:linear-gradient(284deg, rgba(39,43,72,1) 0%, rgba(33,47,183,1) 100%);color:transparent;-webkit-background-clip:text;background-clip:text;}Gyanachand Panigrahi's Linkedin Analytics

Gyanachand Panigrahi

What is Gyanachand talking about?

Gyanachand Panigrahi's Best Posts .css-qqfgvy{font-size:var(--chakra-fontSizes-md);}(last 30 days)

Want to drive more opportunities from LinkedIn?

Gyanachand Panigrahi's Linkedin Analytics

Gyanachand Panigrahi's Best Posts (last 30 days)