Phani Srikanth

Introducing DynaRoute - intelligent model routing for efficient LLM usage At Vizuara, we (Pritam Kudale, Pranavodhayam P R, Raj, Rajat) have spent the last few months studying how teams use large language models in production. What we found was simple but eye-opening. Most teams rely on one large and expensive model for every kind of query, whether it is a quick factual question or a complex reasoning task. This works, but it wastes a lot of money and computing power. That is what led us to build DynaRoute (https://lnkd.in/dGSx2KRq), an intelligent routing layer that automatically analyzes your queries and sends them to the most suitable model based on complexity, cost, and performance. You do not have to change your application or rebuild your backend. You can continue using your familiar OpenAI-style API, just by switching the base URL and key. From that point, DynaRoute handles everything intelligently behind the scenes. In our internal benchmarks, DynaRoute delivered remarkable results. Real-world employee questions achieved up to 92 percent cost savings, while code-related queries saw around 74 percent savings. All of this happened while maintaining nearly the same quality and accuracy as top-tier models. For teams that are running customer support systems, coding assistants, agents, or RAG-based pipelines, DynaRoute can reduce your inference bills drastically while preserving the quality your users expect. You can think of it as a smart, adaptive router that continuously learns how to choose between models like GPT-4.1-nano, Gemini 2.5 Flash, and Claude Sonnet, depending on the query. We also provide API access so you can integrate DynaRoute directly into your own platform or agent framework. The API is OpenAI-compatible and can be configured with just a few lines of code. Here are our current plans: Free Tier: ₹0 per month, includes up to 50,000 tokens, OpenAI-compatible endpoint, and access to premium models. Plus Tier: ₹888.30 per month, includes up to 3 million tokens, unlimited requests, access to premium models, and priority support. Enterprise Tier: completely custom implementation for companies that need dedicated support, detailed analytics, and on-premise deployment. We also provide custom routing models for enterprise-level optimization. We are currently offering Enterprise Plan API access to selected partners who wish to integrate DynaRoute into their platforms. If you want to optimize the cost and performance of your LLM-based applications without compromising output quality, give DynaRoute a try. 👉 Learn more and get your API key at https://lnkd.in/gTrs9NWA

84

3 Comments

Lately there hasn’t been a dull moment in the R-world when it comes to working with LLMs. And as I said in one of my earlier posts: packages emerge faster than you can type 𝚒𝚗𝚜𝚝𝚊𝚕𝚕.𝚙𝚊𝚌𝚔𝚊𝚐𝚎𝚜() 😬. But the same can be said for updates to those packages! In today’s spotlight: 𝗲𝗹𝗹𝗺𝗲𝗿 version 0.3.0 🐘 If you’re not familiar with 𝗲𝗹𝗹𝗺𝗲𝗿 yet, here’s a short summary: 𝗲𝗹𝗹𝗺𝗲𝗿 helps you to call LLM APIs from R in a consistent way, and of course, this opens up a world of possibilities. It’s proving to be a handy back-end for various other LLM-related packages. You could say it’s the workhorse in R’s “LLMverse”. So what’s new? 🔹a handy new function called 𝚌𝚑𝚊𝚝() which can be used instead of specific provider functions like 𝚌𝚑𝚊𝚝_𝚊𝚗𝚝𝚑𝚛𝚘𝚙𝚑𝚒𝚌() or 𝚌𝚑𝚊𝚝_𝚘𝚙𝚎𝚗𝚊𝚒(). It simplifies the chat interface: you now just specify the provider and model in a single string. 🔹defining tools for function calling is also simplified. A tool is just an extra ability an LLM can use. You’re basically saying: “here’s a function you’re allowed to use, this is what it’s called, what it does, and what kind of input it needs”. Specifying that is more intuitive now. Note that this change is breaking. 🔹we hate big unexpected bills so we like to track costs 💸. The latest version of ellmer contains another pricing data source so there’s info for more providers/models. Simply call the 𝚐𝚎𝚝_𝚌𝚘𝚜𝚝() method. PS: 𝗲𝗹𝗹𝗺𝗲𝗿 and 𝗰𝗵𝗮𝘁𝗹𝗮𝘀 (the Python equivalent of 𝗲𝗹𝗹𝗺𝗲𝗿) are closely aligned. I expect (or hope) 𝗰𝗵𝗮𝘁𝗹𝗮𝘀 will follow the simplified chat interface too. Tool spec is already intuitive there and the cost thing is taken care of in the dev version.

165

18 Comments

Quix was conceived in Formula 1 🏎️ It was back in 2017 when me and Tomáš Neubauer joined forces to run Project MIRACLE. At the time McLaren Racing were moving from desktop-based technologies towards a centralised data platform. It was early in the journey, but we had an emerging Telemetry Analytics architecture and wanted to see how far we could push it. It started with hackathons (the video) - as most interesting things do - but led to showcasing an unrivalled data platform that gave us the capability to: • Ingest and enrich live telemetry data. • Process complex analytical models in real-time. • Deliver insights immediately to Performance Engineers. • Centralise all data, scripts & insights globally. It was impressive, but what actually kick started Quix was the 𝐟𝐫𝐢𝐜𝐭𝐢𝐨𝐧 we had in developing the applications. We ended up running a six-month project with a team of 10 engineers to build specific Telemetry Analytics use cases. And we came across friction 𝐞𝐯𝐞𝐫𝐲 𝐬𝐢𝐧𝐠𝐥𝐞 𝐝𝐚𝐲, including: ❌ Getting live telemetry data into a format the Engineers could use. Raw telemetry is not analytics ready. Our engineers wanted a simple way to 'just' open the data and analyse it in their favourite tools. We typically sent them .csv files 😅 ❌ Getting access to Cloud resources from IT. Standing up the initial infrastructure took a month, and then more delays everytime we needed new services - like a database - or some more VM's 🤯 ❌ Packaging MathWorks #matlab and #simulink models to work in containers There are at least 3 ways to package a model artifact and run it in the cloud. You also have to wire-up the input parameters to your telemetry data channels and test them before you run a test or go racing. The teams working on the models couldn't do it so it went in the Software Engineers backlog ⏳ ❌ Collaborating across teams There are a lot of moving parts in an R&D team. Software Engineers were building data ingestion and normalisation pipelines, Data Scientists were training and backtesting AI models, Performance Engineers were experimenting with parameterized Matlab & Simulink models. We regularly got into a mess which slowed us down 🐢 This was the inspiration for Quix - how might we remove friction to deliver complex business use cases more quixly. 8 years later, we have built >60 features that let engineers focus 💯 percent of their time building solutions.

75

5 Comments

🚀 Excited to share insights from Machine Learning Systems by Prof. Vijay Janapa Reddi, a comprehensive guide that explores the principles, practices, and real-world impact of AI engineering. From the evolution of AI 🤖 to cloud, edge, and tiny ML 🌐📱, this work dives into training, deployment, efficiency, and responsible AI 🌱 while highlighting challenges and opportunities ahead. A must-read for engineers, researchers, and innovators shaping the future of intelligent systems. 🌍💡 #MachineLearning #AI #MLOps #ResponsibleAI #Innovation

94

1 Comment

Generative AI Professional Bootcamp. In less than 3 years, I went from a complete beginner to finishing a PhD, having 400+ citations and getting multiple job offers in machine learning. After returning back to India with my PhD, I started teaching Large Language Models (LLMs) and Generative AI. The “Build LLM from Scratch” series which we started has more than 300000+ views now. It has helped a huge number of students, working professionals and researchers to transition to Generative AI and LLMs. Over the last 6 months, we have received the following 2 questions from working professionals, researchers and students: (1) How can I apply LLMs practically in my job? Can you teach me to work on industrial GenAI/LLM projects? (2) How can I work on LLM research problems and publish a research paper? I realised that there is a gap which needs to be addressed. To address this gap: me, Rajat Dandekar and Sreedath Panat launched the Generative AI Professional Bootcamp 3 months back. Here are the results so far: (1) 10 research papers under progress - all authored by industry professionals and student researchers diving into this field for the first time (2) 5 papers to be sent to NeurIPS 2025 Workshops - the best ML conference in the world (3) All the people in the bootcamp have completed 11 hands on projects which they have added to their resume. The topics cover chatbots, agents, MCP systems and many more! The next batch of this bootcamp starts on 5 September and we still have a few seats remaining! Register here: https://lnkd.in/gnFn6Qq5

49

1 Comment

From concept to implementation to diagnostics — the DiD series is now complete! 🎉 Over the past few weeks, I been sharing a blog series with Banani Mohapatra breaking down causal inference methods - from intuition to methodology. We started with the intuition and assumptions, then moved to a step-by-step implementation guide, followed by a real-world case study on free shipping, and finally wrapped up with diagnostics and robustness checks. If you're a data scientist working with product or growth teams, this is a foundational tool for your causal inference toolkit. Catch up on the full series here and subscribe to the newsletter: https://lnkd.in/ethamg6R If you are curious to learn more from Banani and me, sign up for our upcoming AB Testing workshop: https://lnkd.in/gpxBwvrM

40

3 Comments

🚨 India’s consumption puzzle: Why do two official datasets tell different stories about how we spend? PFCE (top-down; macro-consumption) vs. HCES (bottom-up; micro-consumption) 🔹 Gap has widened from 5% in 1972–73 to ~45% today. 🔹 Free/subsidised services (education, healthcare, meals) inflate PFCE but barely appear in HCES 🔹 Food share in household budgets is falling; non-food (transport, healthcare, rent, education) dominates This divergence isn’t just statistical—it affects how we measure poverty, prosperity, and the impact of policy. 📊 Latest piece by Ashish Kumar and Payal Seth from Pahlé India Foundation (PIF) in businessline unpacks the data divide and calls for urgent reforms (https://lnkd.in/dJWjyCQe)

179

8 Comments

Recently, I came across news about PhysicsWallah’s new AI model that beat GPT-4 on JEE math questions. Naturally, the internet split in two—some proudly calling it India’s moment, others waving the “it’s just fine-tuning” flag. Here’s the reality: This is a small language model, not a large one. And its job wasn’t to write novels or generate Python—it was built specifically to excel at JEE math. Which it did, remarkably. But here’s where many get it wrong. Just because a domain-specific model outperforms GPT-4 on one task doesn’t make it more “intelligent.” That’s like saying a calculator is better than a scientist because it computes faster. So I wrote this piece to help business leaders, investors, and AI enthusiasts avoid falling for misleading headlines like “We built a model better than GPT-4 at X.” It probably is. But only at X. The article breaks down: • The difference in training goals, data, and infrastructure • Why smaller models are cheaper, safer, and sometimes better—but not general • How to spot hype vs actual product strategy Give it a read below if you’ve ever caught yourself thinking “why not just build our own GPT?” https://lnkd.in/gT-_JWnZ #AI #LLM #SmallLanguageModels #GPT4 #EdTech #DeepTech #AIProduct #MachineLearning #DomainSpecificAI #VC #AIForBusiness #LinkedInAI #AIethics #AIgovernance #LLMvSLM

30

1 Comment

I have covered 2 research papers on RAG (Retrieval Augmented Generation) in my recent youtube videos. 🚀 1. Self RAG [published Oct 2023] ✅ 2. Corrective RAG (C-RAG) [published Oct 2024] ✅ In the videos, we implemented the above papers from scratch in LangChain and Python and also did a detailed Code walkthrough. If you're interested to know how to build them, feel free to give them a watch! 💯 These are a part of my "AI Engineering 101" playlist. ❤️ Link to the channel is in the comments. 🔗 #aiengineer #youtube #rag #programming #coding

114

8 Comments

India has the minds. Over 2.5 million STEM grads a year (AICTE, MEITY). AI papers from IITs and IISc making it to NeurIPS and Nature. Engineers leading AI divisions at Google, Microsoft, OpenAI. Yet… no trillion-dollar deep-tech giant born here. It’s easy to blame infrastructure. But the bigger reason might be cultural, we don’t celebrate unfinished brilliance. We’ve repeated the “infra, funding, policy” argument so often that it’s become noise. The less obvious truth is India has a low tolerance for visible failure and that kills deep-tech at birth. As a nation we don't bet on high risk tech, deep tech, and long term fruits. We are always looking at low hanging fruits, we expect same from our entrepreneurs. Look at Nvidia... their first GPU prototypes in the mid-90s overheated, failed under load, and couldn’t even demo reliably. Yet, Jensen Huang convinced investors to double down because they saw an acceleration curve, not a polished product. That willingness to back trajectory over perfection is what allowed Nvidia to ship, stumble, iterate, and dominate. In India, that curve flattens early. A half-baked product that fails in public isn’t a learning milestone, it’s a reputational stain. Founders fear it, investors penalize it, babus politicize it. customers walk away. So radical prototypes stay in labs until they’re “perfect” By which time the market has moved on, or a global player has already claimed the space. DeepSeek’s story isn’t just about technical genius; it’s about harnessing speed and iteration in full view of the market that encouraged it. Until India shifts from risk aversion to risk metabolism, treating early stumbles as assets rather than liabilities ,we’ll risk exporting our most ambitious ideas to ecosystems that know how to grow them in the wild. And when they finally succeed elsewhere, we’ll call it a brain drain problem, when it was actually a permission-to-fail problem all along. #StartupStories #GuerrillaMarketing #Leadership #MouthShut #AutoRickshawAds #FoundersDiaries #IndiaInternetStory

72

20 Comments