Avani Bajaj

Wanting to learn about enterprise AI agents but not sure where to start? I highly recommend Subha Shetty's Mastering Enterprise AI Agents course. I met Subha through Sidebar last summer and was immediately impressed by her depth of knowledge. She's both a practitioner (Chief AI Officer at Wayfound) and a natural teacher who is genuinely invested in her students' success. The course gives you a strong foundation across everything that matters: - Understanding enterprise agents and when to introduce them - Evaluating implementation costs and calculating ROI - Navigating governance and compliance considerations - Hands-on practice actually building an agent and thinking through an enterprise case study If you're serious about getting up to speed on enterprise AI, this is a great place to start.  Her next course starts on May 18th.   Don’t wait to sign up.  She keeps each cohort small and it fills up quickly! 🔗 https://lnkd.in/e-PW9tqc Have you taken any AI courses worth recommending? If so, drop them in the comments.

6

Teaching LLMs to be good at writing has been a challenge. Over time, we have developed a knack for calling our AI generated content. This is because the LLMs fall back to same pitfalls and have known patters, which often feel too ‘scientific’ for lack of a better word, lacking in nuance and emotion/experience of human beings. An interesting article in Atlantic on how labs are trying to solve this by the brilliant Jasmine Sun. #AI #LLMs #ArtificialIntelligence

9

1 Comment

Ten Years，Three Lessons Today marks my 10th year at Amazon. When I started my career, I never imagined staying at the same company for a decade. But Amazon is not just any company - it's a place that constantly reinvents itself, challenges you to think big, and reward builders. A friend today reminded me: “You only get so many 10 years in your life.” That struck a chord and prompted me to reflect—not just for myself, but also for the younger version of me and the early-career professionals who’ve reached out for advice. I’ll keep it short and share three lessons that have shaped my journey: 1. Choice is more important than effort. I used to believe that hard work alone guarantees success. Over time, I learned that what you choose—your team, your manager, your priorities—can matter even more. Effort is necessary, but making the right choices gives that effort leverage. As we gain experience, we gain the ability—and responsibility—to choose better. 2. Invest in relationships. One of the most important reasons I’ve stayed at Amazon for 10 years is the people. I’ve had the privilege of working with brilliant, generous colleagues who’ve taught me more than any textbook ever could. Some of these colleagues have become mentors, collaborators, and even dear friends. Investing in relationships built on trust, respect, and kindness has made all the difference. 3. You can’t have it all—at least not all at once. Every choice involves trade-offs. Especially as a woman, I’ve had to navigate moments when I couldn’t be fully present at home and pursue every professional opportunity. And that’s okay. We are defined not by what we have, but by what we choose. Looking back, I wish I had taken more risks. I held myself back from opportunities I thought I couldn’t succeed in, without even trying. So for the next 10 years, my goal is simple: be bold, be brave, and choose.

1,021

35 Comments

Great collaboration with you Amauri and Ovidiu-Constantin on voice enable control of a robotics arm!!! Looking forward to further extensions to this work!! PhysicalAI has entered an exciting phase with VLM, LLM, VLA and World Models!! 🦾 🚀 #PhysicalAI #Robotics #NovaSonic #VLM #VLA #AmazonNova #AWS

1

🚀 Target Multiplai 2025: AI with Purpose, Scale, and Empathy Just returned from an incredible day at Target’s AI Conference in Bengaluru, and I’m truly inspired by how AI is shaping real-world impact across industries and communities. 🔹 Agentic AI Is Evolving Brad Thompson (SVP, Target Tech) shared frameworks for scaling multi-agent systems, highlighting the importance of balancing autonomy with centralized control—making AI more responsive and practical for enterprise use. 🔹 AI at Scale for a Billion+ Jigar Halani (NVIDIA) showcased India’s use of real-time AI for citizen services—addressing ethics, personalization, and ultra-low-latency decision-making for massive populations. 🔹 Empathy-Driven Innovation Manu Chopra (Karya) delivered a powerful reminder that inclusive AI can create real economic impact. His work is enabling rural communities through AI-powered job opportunities and building more representative datasets. 🔹 Retail’s AI Transformation Target’s real-time pricing engine, built by Rohit Punnoose & Shamally Chopra, is a game-changer—optimizing billions of pricing decisions by aligning competitiveness, customer value, and business growth. ✨ From architecture to accessibility, every session reflected one truth: AI’s future is not just intelligent—it’s inclusive, intentional, and human-centered. #AI #TargetMultiplai2025 #AgenticAI #EthicalAI #RetailTech #InclusiveInnovation #Databricks #NVIDIA #Karya #RealTimeAI #DataScience #MachineLearning

8

1 Comment

I’m excited to share a project I’ve been building: an Autonomous AI Research Agent that runs quietly in your system tray and does the heavy lifting for web research. 🤖🚀 I wanted a tool that didn't just give me a quick chatbot answer but actually planned out a research strategy, browsed the web, read RSS feeds, and synthesized it all into a clean, downloadable Markdown report. So, I built it. Here is what’s under the hood: 🧠 Multi-Step Reasoning: Powered by LangChain, the agent breaks your prompt down into a structured execution plan before it starts searching. 🌐 Dynamic Tools: It autonomously decides when to use DuckDuckGo, scrape specific URLs, or read RSS feeds to gather context. 🔒 Privacy-First Option: You can plug in API keys for OpenAI/Gemini/Anthropic, OR you can run it entirely offline on your own hardware using LM Studio. 🖥️ Zero Setup: I built the UI with Gradio and packaged it all into a standalone .exe for Windows. No Python environments needed for the end-user! If you are interested in AI agents, local LLMs, or just want a cool research assistant on your desktop, you can check out the source code or download the .exe directly from my GitHub: 🔗 https://lnkd.in/gQ2hdqG9 I'd love to hear your thoughts or feedback! What features should I add to the agent next? #AI #Python #LangChain #OpenSource #LocalLLM #LMStudio #Gradio #SoftwareDevelopment #GenerativeAI

3

A few grounded takeaways on building effective agents in the real world from the panel discussion with Suba Karuppan and Pankaj Kamble at the Generative AI Summit DC, 2026 hosted by AI Accelerator Institute.    - Memory is a crucial component of an agent's cortex. Distilling the right facts and avoiding memory overload is foundational to reliable agent behavior. - Context is a resource with diminishing marginal returns. Keeping it finite and using it wisely is essential to balancing accuracy, latency, and cost in production systems. - Hype is ahead of reality. Multi agent demos move at 100 mph while real world systems are closer to 10 mph. Engineering constraints still apply. - Production readiness matters. Observability, auditability, security, and reliability are required for systems to work safely at scale. - Humans remain in the loop. For sensitive or high accuracy workflows, full automation is still the exception. Conclusion: Real Agentic AI success comes from well designed systems. Full length panel discussion here: https://lnkd.in/gRX5TzKD What are some of your learnings from building agentic systems in real world? Please do share in the comments. . . #AgenticAI #PanelOnAgents #AIEducation #AIStrategy #AIForLeaders #AIforPractioners #GenAI #Langgraph #Moltbot #Claude

40

Wow, this one took longer than I expected however continuing on path for learning more about LLMs and RAG, here is an article where I have listed some key learnings for anyone trying to understand what LangGraph is and key components of execution! #langchain #langgraph #agenticrag #rag #llm

17

You are attending a Gen AI Interview with Amazon. Interviewer: Walk me through the tradeoffs between on-index (precomputed) vs real-time retrieval for user queries in a RAG product, when would you pick which? Explanation: 1) On-index (precomputed retrieval) In this approach, all documents are preprocessed offline, chunked, embedded, and stored in a vector index. So at query time, we only do embedding + vector search, which is very fast. The big advantage is performance and scale. Search usually takes tens of milliseconds, which makes it ideal for products serving thousands or even 100k concurrent users. It’s also cheaper because embeddings are computed once, not per request. The downside is staleness. If the data changes frequently, we need re-indexing, which is expensive and delayed. Also, it’s not great for user-specific or transactional data. So I prefer on-index when: data is mostly static like policies, docs, FAQs, knowledge bases and we care about low latency and high concurrency 2) Real-time retrieval Here, instead of relying only on a prebuilt index, we fetch data live at query time, maybe from a database, API, or tool. So it’s always fresh and personalized, which is great for things like: order status account balance tickets user history But the tradeoff is latency and reliability. Every request depends on external systems, so response times become unpredictable, and it’s harder to scale. Also, cost per query is higher. So I use real time retrieval when: freshness is critical data changes frequently or answers must be user-specific ➡️How I design it in production In practice, I rarely choose only one. I use a hybrid approach. Static knowledge --> on-index Dynamic/personal data --> real-time tools For example: If a user asks “Why was my refund rejected?” i would: retrieve refund policy from vector DB fetch that user’s refund record from DB then let the LLM combine both That gives: fast retrieval fresh data and better accuracy #rag #llm #genai

84

15 Comments

Use Neo4j’s query language Cypher to manage and retrieve data stored in knowledge graphs. Write knowledge graph queries that find and format text data to provide more relevant context to LLMs for Retrieval Augmented Generation. Build a question-answering system using Neo4j and LangChain to chat with a knowledge graph of structured text documents.

4

The original intent of starting this podcast was a forum to teach. But within a few episodes, I realized it was actually a much higher-leverage resource for accelerating my own learning. In today’s episode of the Data Neighbor podcast, our guests are taking the day off. Hai, Sravya, and I use the time to look back on what exactly we were thinking when we started this thing a year ago, everything we’ve learned along the way, and how this podcast has connected us to so many people we never would’ve met otherwise. We talk about the stuff we did not expect at all (both the good parts and the parts that are way more work than they look from the outside). Honestly, this one is just a really fun episode. Only downside: you have to listen to me ramble a little more than the pleasure of our much more well-spoken guests. Still a good time, and a nice way to wrap up the year. omg is this Data Neighbor Wrapped?! Check it out in the comments ;)

6

1 Comment

Over the past couple of years, I’ve worked extensively with large language models. In machine learning, it’s often said that nearly 80% of the effort goes into curating high-quality data, with the remaining 20% spent on training the model. That has largely been my experience as well. I’ve spent countless hours curating datasets and training models so they could perform well on internal benchmarks. After the hardest parts - curating the data and training the model are behind you, it’s tempting to consider the work mostly done. In practice, however, this is only the beginning. Models rarely improve in isolation. Meaningful progress depends on real user feedback, and that requires making the model accessible to a large number of users. Deploying a model at scale introduces an entirely different set of challenges, ones that are often more complex than training the model itself. Beyond state-of-the-art architectures and modelling techniques, what truly enables adoption is a reliable and scalable system around the model. When building such systems, we have to keep two major constraints in mind - - Users should not have to wait long for a response - Overall latency must remain low enough for the experience to feel seamless. I have explored these challenges in a two-part article series, focusing on the system-level considerations required to serve LLMs reliably and efficiently at scale. Below, I’m sharing the first article in the series. https://lnkd.in/gwNah5uu Thanks for taking out time to read this, and stay tuned for the second article which will be dropping soon!

89

2 Comments

Over the last few days, I dove deep into how LLMs understand meaning, not just words through the "LLMs and Semantic Search" course by DeepLearning.AI. This course helped me connect the dots between: 1. Converting input into vector embeddings 2. Using those embeddings for semantic retrieval 3. Applying RAG (Retrieval-Augmented Generation) to make LLMs context-aware 4. Designing smarter data pipelines for unstructured information Here’s a simple flow of the process that really stuck with me: [User Query]    | [Natural Language Processing]    | [Embedding Model]    | [Vector Embeddings]    | [Similarity Search / Vector DB]    | [Relevant Context Retrieved]    | [LLM + RAG → Final Answer] Gained a lot of clarity on how these systems actually work — and now it’s time to put it into practice. #SemanticSearch #LLMs #RAG #NLP #AI #DeepLearningAI #MLWorkflow

25

I benchmarked 18 forecasting models — from LightGBM to foundation AI. A 1970s algorithm nearly won. The score gap was 0.03. The operational gap was enormous. https://lnkd.in/dTfRUxrB

12

If you are someone pushing boundaries of LLM interaction, improving prompt engineering strategies, and most importantly building more robust AI application, this paper on agentic context engineering could be a good read for you !! https://lnkd.in/gG97xZGk

40

1 Comment

Reflecting on Target Multiplai 2025: Where AI Meets Real-World Impact Had an inspiring day at Target’s AI conference in Bengaluru! Here are my top takeaways: 🔹 Agentic AI is Here to Stay Brad Thompson (SVP, Target Tech) emphasized pragmatic frameworks for scaling multi-agent systems, tackling integration challenges while balancing autonomy and control. 🔹 Real-Time AI for Billions Jigar Halani (NVIDIA) showcased how India leverages AI for hyper-personalized citizen services—navigating ethics, scalability, and real-time decision-making at unprecedented scale. 🔹 Inclusive AI Drives Systemic Change Manu Chopra (Karya) highlighted empathy-driven design—using AI to create economic pathways for rural communities while ensuring datasets represent diverse voices. 🔹 Retail’s AI Revolution Target’s real-time pricing engine (powered by Rohit Punnoose & Shamally Chopra) demonstrated how AI balances competitiveness, profit, and customer value across billions of dynamic decisions. Grateful for the deep dives into ethical scaling, emerging architectures, and AI’s power to uplift. The future isn’t just automated—it’s intelligent, inclusive, and intentional. 🙌 #TargetMultiplai2025 #AI #GenerativeAI #MachineLearning #TechInnovation

77

1 Comment

🔍 "RAG Won’t Save You — Unless You Understand NLP First." Let’s talk about what’s really behind the hype around Retrieval-Augmented Generation (RAG). Everyone’s building “RAG apps” today — but very few actually understand what makes them work. Here’s the hard truth: RAG isn’t about large language models (LLMs). It’s about retrieval engineering — classic NLP meets modern AI infrastructure. 🧠 The Core Idea RAG = Retrieval + Generation But the “generation” part — the LLM response — is the last step in the pipeline. The real magic lies in what happens before that. If your RAG system isn’t performing well, 9 out of 10 times the issue is in your retrieval pipeline, not your LLM. ⚙️ The Engineering Layer Nobody Talks About To build a solid RAG system, you need to nail each layer: 1. Chunking Strategy:     Don’t split text blindly. Use semantic or recursive chunking to preserve context.     (Cutting mid-sentence = noisy embeddings = garbage retrieval.) 2. Embedding Models:     Choose wisely — sentence-transformers, multilingual, or domain-specific models can drastically affect recall and precision. 3. Similarity Search:     Cosine vs. dot-product vs. dense reranking — each has trade-offs in latency and accuracy. 4. Vector Indexing:     Use the right vector DB — FAISS, Chroma, Weaviate, or Milvus — and tune your ANN (Approx. Nearest Neighbor) parameters. 5. Context Stitching:     Injecting retrieved chunks into prompts isn’t trivial.     Context windows, token limits, and ordering logic matter a lot. 🚫 It’s Not “Just Add Context to OpenAI” A well-engineered RAG pipeline handles: - Low-relevance edge cases via reranking - Hallucination reduction via better retrieval control - Component chaining (Retriever → Reranker → Prompt Template → LLM) That’s real AI engineering, not prompt hacking. 🎯 The Takeaway RAG is 80% retrieval and 20% generation. The LLM is the cherry — but the cake is built from solid NLP fundamentals. If you can design that foundation, your Agentic AI systems won’t just respond — they’ll reason. #RAG #AIEngineering #AgenticAI #NLP #LLM #VectorDB #GenAI #LangChain #ChromaDB #MachineLearning #MLEngineering

13