Raja Appachi’s Post

More news on the decision science front! Google are now offering algorithms for solving the liner shipping network design problem as a SaaS (solver-as-a-service)⛴️   It was not many weeks ago that Nvidia announced their cuOpt (SaaS) for solving routing problems, and now Google with a new announcement. Google has being working a long time in the decision science area internally and externally by e.g. releasing a good Python package for solving this kind of problems called ORTOOLS (although the shipped MIP solver of the package is actually SCIP – which is not developed by Google)😮   This SaaS is very industry specific, and I know the problem has been heavily studied in both literature and by companies such as A.P. Moller - Maersk. The most exciting part is perhaps the last paragraph of the blog post which is a cliffhanger😦   “The Shipping Network Design API is one of a set of Operations Research APIs that we'll be adding to in the future as we look for other industries to optimize.”😱

🔥 XGBoost, which stands for Extreme Gradient Boosting, is an #OpenSource, scalable, distributed gradient-boosted decision tree (GBDT) #MachineLearning library. Although XGBoost isn't new, it remains a popular favorite for supervised #ML tasks like regression, classification, and ranking. XGBoost algorithms are designed, in particular, to be highly efficient, flexible and portable. And XGBoost is seamlessly integrated with Comet! 👇 Try the Colab to get started: https://lnkd.in/exsuUReS #MLOps #DataScience #Integration

Do you want to build a simple chatbot using the OpenAI API? I've created a notebook that will guide you through the process in just a few minutes. You can access it https://lnkd.in/g8wcNPyS If you prefer to run it as a project, you can find the complete setup and code on my github : https://lnkd.in/g_Zg-WG3 #chatbot #openai #genai

3MTT KNOWLEDGE SHOWCASE NAME: JOHN UDEME DAVID FELLOW ID: FE/23/9724901 TRACK: DATA SCIENCE September Knowledge Showcase: Career Roadmap (2024–2029) In the next five years, I envision myself becoming a Data Scientist and eventually transitioning into a leadership role as a Lead Data Scientist or Machine Learning Engineer. I aim to leverage data science, machine learning, and networking skills to solve business problems across industries, focusing on scalable AI solutions and data-driven decision-making. 1. Steps to Achieve Career Goals Year 1 (2024-2025): Build Technical Expertise & Certifications Skills Mastery: Focus on Python machine learning, data analysis tools (e.g., Tableau, SPSS, IBM Watson), and cloud computing (e.g., AWS, Google Cloud). Certifications: Complete CISCO Network Security, Google Data Engineer, and AWS Machine Learning certifications to solidify my data science and cloud skills. Projects & Portfolio: Develop personal and open-source projects, like text and video annotation tools using Python and third-party software, and create predictive models for real-world datasets. Year 2-3 (2025-2026): Deepen Specialization & Networking Advanced Projects: Lead machine learning projects that apply AI to fields like healthcare or finance, incorporating tools like Atlas.ti and Survey CTO. 3MTT Fellowship Projects: Complete the 3MTT Knowledge Projects (such as text and video annotation models), adding them to my portfolio for industry visibility. Networking & Collaboration: Attend data science and AI conferences, publish findings on platforms like Medium or Kaggle, and collaborate with global experts via GitHub and LinkedIn. Year 4-5 (2027-2029): Leadership & Mentorship Lead Teams: Take on a leadership role, managing data science teams or leading the data division of a tech firm. Mentorship: Give back by mentoring upcoming data scientists and networking engineers through platforms like CISCO Academy or local mentorship programs. 2. Mentors & Inspirations Mentors: I look up to leaders like Andrew Ng, the co-founder of Coursera, for his work in AI education. Locally, I seek guidance from Dr. Ifeoma Ikenna, a senior data scientist at NITDA, and CISCO Academy networks. 3. Challenges & Milestones Challenges: Balancing my learning with my current Data Analyst and Network Engineer role may be tough. Overcoming imposter syndrome in the fast-paced tech industry is another challenge. Milestones: My success will be marked by reaching key milestones, such as securing leadership roles, publishing industry-relevant papers, and obtaining high-level certifications. PROJECT TITLE: Analyzing Customer Churn for a Telecom Company Using Exploratory Data Analysis (EDA). Link to the mini–project @: https://lnkd.in/esQYq8XX #My3MTT #3MTTWeeklyReflection Tag: @3mttNigeria, @IHSTowers, @NITDANigeria

🤖 Here is a 10-minute quick start Colab on LightRAG. It includes: 1. A simple chatbot with OpenAI and Groq. 2. A RAG pipeline with batch embedder and FAISS semantic retriever. https://lnkd.in/g23YDaHt Feedback is appreciated as always. #lightrag #artificialintelligence #machinelearning ______________________________________________ LightRAG : The PyTorch Library for LLM Applications. It is light and modular, with a 100% readable codebase. Follow + hit 🔔 to stay updated.

Recently, I've been exploring ways to enhance RAG for CSV data, and proofing why blindly trusting orchestrators to process your data in the most optimum way for your use case is not a good idea. A key insight: removing column names from the content for embeddings can significantly improve retrieval accuracy and vector distinction. Key findings: • Increased cosine similarities between query and relevant records • Improved ranking of results • Greater average distance between vectors This seemingly small change led to a notable boost in Signal-to-Noise Ratio, enhancing the overall performance of the RAG system. It's a reminder that in AI/ML, application architecture choices matter immensely. While orchestrator abstractions are convenient, diving deeper into data preprocessing can unlock substantial performance gains. What's your take on balancing convenience with performance in AI systems? Have you encountered similar optimizations in your work? Check out my notebook here -> https://lnkd.in/dbqCAcdW #AIEngineering #MachineLearning #DataScience #RAG #llm #openai #langchain

MLops Week 1: This week, I got introduced to Machine Learning operations. The data used for this project is the New York taxi-rides' 2023-monthly data. The January 2023 data was preprocessed and trained with a plain linear regression model, while the February 2023 data was used to predict the expected time arrival of a trip -- in minutes. Furthermore, the model was evaluated with the root-mean-squared-error metric, to determine how well the model performed. Special thanks to Alexey Grigorev and his team. I look forward to learning more MLops technological tools! #MLops #DataTalksclub Attached below is the project for perusal: https://lnkd.in/gwXmdYAY

Investing any time to learn #tensorflow is a very bad idea. I've been using it since 2016-2017 and i have come to regret it. The software in itself is of extreme complexity, lightly wrapped up by keras. The support from #google is minimal, documentation is automated and outdated. The software consists of towers of abstractions and builders upon builders. The bugs endless, catastrophic and without any reasoning. The changes constant and backwards breaking. It is also extremely wasteful with resources (memory to be exact). The CUDA <-> Tensorflow tug of war exhausting. Case in point: I wanted to use convnext for a project, very nice there is an already pretrained version. So I pick that up, make my changes, try to run it, ok it works, try to checkpoint it => gianormous stack trace. Following that up, a simple not-naming a weight variable from convnext is breaking everything. (Edit: This implies nobody tested it before release which is extremely bad).Hot swapping a custom layer with the fix in solved the issue but wasted 3 hours not working on the problem but working on side stupid issues. In addition to that #google is constantly flirting with this new jax framework which is supposed to be small and nimble and sexy. I wont be jumping on that any time soon. Unless you are running some kind of ultra-large cluster i see very little reason to be onboard with google on anything. On the other hand pytorch has a much better community, tooling and documentation. I really hope decoupling keras out of it brings back some sanity but i highly doubt it. Google is not producing great software and from an outsiders perspective this project has no direction. > Venting out. #artificialintelligence #aiengineers #machinelearning #deeplearning #deepneuralnetworks

I tested Anthropic's new prompt caching. First of all - what is it and why do you need it? My results are the following: 🕐 Saves a part of your prompt for faster and cheaper re-use. 🕑 It's for larger queries that either have loads of examples, a large corpus of text or in general a lot of repetitive content. 🕒 Response times are improved 30-90%! 🕓 Costs are decreased 50-90%! Secondly - let's dive deeper. 🤔 I right away spot the elephant in the room. Creating the cached content is over 20% more expensive, but using it is almost 10% cost. All good, right? No. The cache is active for only 5 minutes, so if your prompts are not ran every 5 minutes, you will pay a premium and won't be able to get the benefits of the reduced costs and response times. 🤨Wait, if the prompt is the same and the answer is the same then why do I even use an LLM call? Good question! And the answer is that you can still have variance within the cached prompt, it will only "save" the prefix, so you can still have client input or other variables IN the prompt. Anthropic has solved this issue in an interesting way! 😱IMAGES! You can cache images and use computer vision with this new feature! I can think tons of applications where this can come in handy, when you need to frequently prompt the same image(s) over again. 🧐 The cache might be emptied in other conditions other than the 5 minute cap. If your message(s) content changes too much, the cache will be invalidated. This means tool usage and the order of messages within the API call. Thirdly, let's test it out. Google Colab https://lnkd.in/d2e8ZMnu Price difference vs OpenAI OpenAI GPT-4o Pricing: Input: $5 / 1M tokens Claude Sonnet 3.5 Pricing: Input $3 / 1M tokens, $3.75 for creating cache, $0.30 for using cache Let's assume you have a fixed 100k input tokens, 1k output tokens, you run it 100 times an hour: Claude 3.5 sonnet without caching: $31.5 per hour Claude 3.5 sonnet with caching: $10.95 an hour OpenAI GPT-4o: $51.5 per hour Summary - There are direct benefits to this new approach, but one needs to keep in mind about the limitations and restrictions (especially about the cache lifespan and invalidation rules). - Anthropic provides the cached token creation count, as well as the cached token read count, so you can easily keep track if this feature makes sense to you. - Use it so that you add your repetitive content in the BEGINNING of the prompts/messages. - This is a great cheap alternative to RAG that might even have better performance. - Pay attention to Claude API rate limits, it's very easy to hit them. - The keyword here is REPETITION. If you need to query the same text/data/images many times in a small timeframe, then this is the solution for you! - I did not notice a MAJOR win in time, but certainly a win in COST. Have a great one! <3

LeetCode says 2Sum was asked by FAANG alot in last 6 months. Is that true? With LeetCode premium one can see how often companies ask each question, and 2Sum seems to be one of the most popular ones. It says it got asked by Amazon, Microsoft Google Facebook etc. but it seems weird that they ask such easy question even right now? Does anyone know if it is true or where did it come from? Maybe its only one of questions for OA or something just a warm up question! Join the warrior's Community :- https://lnkd.in/eWcY5TT4 #leetcode #twosum #datastructure #algorithms #dsa #leetcodequestions #interviewquestions #amazon #google #microsoft #facebook