Prashanth M.

🦺 🛟 Agent Safety > Model Safety LLMs didn't break your app. Agents can. As we wire LLMs to tools, browsers, payments & code the risk shifts from bad text to bad actions. Top failure modes (seen in the wild & in papers): 🚩 Prompt/Content Injection -> Tool Misuse. Malicious web pages or data instruct agents to steal secrets or make unintended API calls. This is now a primary risk for browsing and MCP - style tool agents. 🚩 Over permissioned tools & long-lived tokens. Agents get "god mode" scopes. A single injection becomes account takeover or irreversible ops. Security notes for MCP emphasise least privilege and short-lived auth. 🚩 Unsafe web autonomy. Benchmarks show web agents will attempt harmful tasks unless constrained (posting misinformation, illicit sales, etc.). You must measure misuse not just assume guardrails. 🚩 Supply-chain & retrieval poisoning. Agents trust plugins, third-party tools & indexed data that an attacker can taint. OWASP’s newest GenAI Top 10 calls this out explicitly. 🚩 Process safety gaps. NIST's GenAI Profile (AI 600-1) warns that organizations ship agents without role clarity, human-in-the-loop or incident playbooks. ✅ Model safety reduces bad text. Agent safety prevents bad state. Treat the agent like a junior SRE - with narrow permissions, audited actions and clear escalation paths. If you're building agents, what's the one control you won't ship without? Also drop a comment if you'd like me to share a follow-up post on how to tackle each of these risks. #AgenticAI #AISafety #AIAgents #GenAI #CyberSecurity #MCP #OWASP #NIST #AITrust #ResponsibleAI #AIForBusiness

5

Running LLMs on mobile phones—completely offline and locally—has become a reality. I recently tried the Google AI Edge Gallery app and was surprised by how effortlessly it brings generative AI to your device, no internet connection required after the initial setup. The app lets you chat, experiment with prompts, and even analyze images, all on your phone. If you're curious to try on-device GenAI for yourself, here’s the link: https://lnkd.in/dFJ9MXWV #GenAI #EdgeAI #LLM #MobileAI

35

2 Comments

RELEX Solutions AI-driven solutions are helping DoorDash streamline inventory management and optimize fresh product lifecycles across its DashMart network. It’s all about reducing waste, improving accuracy, and ensuring faster, more reliable deliveries to consumers.  Learn more: https://relex.link/47ir1I7  #AISupplyChain #RetailTech #InventoryOptimization #Sustainability  

1

I do not really post a lot about tech stuff, but maybe I can start. Here is a paper I read recently. ROSE: Robust caches for Amazon product search [https://lnkd.in/gPhbgHHQ] Thanks to Arpit Bhayani's paper shelf recommendations. The insights from this are really fascinating for the search world and how Amazon always knows how to find what you are looking for even if you have clumsy fingers like mine. For example if you are looking for shoes, you might type something like: “nike shoe,” “nike shoes,” “nikes shooes.” A normal cache treats these as different, so it gets big and slow, and you miss the cache a lot. Amazon’s ROSE fixes this by caching the intent, not the exact spelling. What ROSE essentially is, is a typo/variant-tolerant cache for product search that groups similar queries into the same “bucket,” so most lookups hit fast without growing memory. It’s deployed in Amazon Search and improved both latency (single-digit ms for most traffic) and business metrics. How this works is even more fascinating. 1. Locality-Sensitive Hashing: Similar queries collide in the same bucket. • Lexical-preserving hashing (character n-grams + minhash) for typos/variants. • Product-type-preserving hashing (weighted minhash) to keep the category intent (e.g., “dishwasher” vs. “dishwasher parts”). 2. Reservoir sampling: caps each bucket so memory stays constant even as queries grow. (Super convenient, and one of the many engineering gotcha's people might miss) 3. Count-based k-selection: avoids pairwise similarity inside buckets; just count collisions across hash tables → near constant-time retrieval. Where we can apply some of these concepts with Gen-AI (which is in many flavors essentially a search problem with extra steps)- Semantic prompt cache: Group paraphrased prompts to reuse an answer/logits/KV cache even when text isn’t identical. Can be helpful in saving your tokens and $. RAG request cache: map similar user questions to one retrieval plan to cut vector searches + reranking cost. [I remember Azure AI search having a RAG cache, but I am not sure how that worked] Tool/agent call cache: Dedupe near-duplicate function/tool invocations (e.g., same API call phrased differently). Eval & feedback loops: Bucket similar generations or errors to reuse critiques/patches. Support/search front-ends to LLMs: Typo-tolerant, intent-stable pre-rewrites before hitting the model, ensure we only use tokens when absolutely needed.

174

6 Comments

Poor AI dependent Customer service- Amazon Amazon I am writing to formally raise a grievance regarding the consistently poor customer service and flawed pickup process I’ve experienced with Amazon’s Sales & Logistics team. Despite multiple attempts to initiate a return, the pickup process has been repeatedly mishandled. The app fails to provide any contact details for the assigned pickup agent, leaving customers completely in the dark. Worse, pickups are frequently cancelled or rescheduled without any prior notice or consent from the customer. This lack of transparency and communication is not only inconvenient—it’s unacceptable. I expect Amazon to uphold a basic standard of accountability and customer respect. Random cancellations and opaque logistics are not just frustrating—they erode trust in your platform. I request the following actions be taken immediately: Share pickup agent contact details via the app or SMS for coordination. Ensure pickups are not cancelled or rescheduled without explicit customer confirmation. Provide a clear escalation path for logistics-related grievances. Please treat this matter with urgency and respond with a resolution timeline. I hope Amazon values its customers enough to fix what is clearly a broken process.

8

3 Comments

Is the "Search & Browse" era of E-commerce officially evolving into something entirely new? 🛒🤖 I just finished watching an incredible deep dive with Manav Garg (Together Fund) on Think School. As a developer in e-commerce space, one takeaway really stood out: The transition to Agentic Commerce. For years, the industry has been obsessed with perfecting filters and recommendation engines. But the next big leap isn't a better UI, it’s the "headless" interaction between AI agents. (Timestamp: 18:05) The Shift: From UI to Infrastructure In the near future, customers may not spend 20 minutes scrolling. Instead, personal AI agents will negotiate with brand agents to find the best deal based on a user's specific goals and budget. (Timestamp: 15:08) What this means for the future of our craft: Logistics is the new Front-end: If the "Browse" layer becomes automated, the real competitive edge moves to the physical layer i.e supply chain, speed, and fulfillment infrastructure. This is where vertically integrated platforms really shine. (Timestamp: 21:31) The Rise of "Orchestration": With tools like Cursor and AI-assisted coding, we're moving away from boilerplate and toward building complex "Orchestration Layers" that allow these agents to talk to our platforms seamlessly. (Timestamp: 08:30) Efficiency as a Feature: As we integrate LLMs, the focus moves to making them profitable through smarter token usage and caching, turning "hype" into sustainable margins. (Timestamp: 36:52) We often focus on the 95% of time users spend browsing. (Timestamp: 17:52) But as that time shrinks, the platforms that win will be the ones built for agents, not just humans. Are we ready for Agentic Commerce, or are we still building for a world that's quickly moving past the search bar? Video Source: https://lnkd.in/ghfmbqw3 #Ecommerce #SoftwareDevelopment #AI #AgenticCommerce #ThinkSchool #Innovation

22

I’ve interviewed to 500+ candidates combined in the last 7 years of my journey at Amazon, Cars24, and now Nielsen…(150+ in the last 1 year alone) This is the best guide I can give you on how to pass technical interviews and what 90% of candidates often forget while interviewing. (Btw, I also used this to switch jobs 3 times in my career) [1] Don’t rush to code, talk through the problem first Many candidates fail because they panic and try to impress with fast coding but the right move is to slow down and talk out loud. Break the problem into conditions, use a whiteboard or pseudocode if needed. Even a simple question becomes easy once you write the logic in plain English. [2] Think before you type Take your time to understand the requirements before you touch the keyboard. Rushing leads to bugs. Thoughtful implementation shows maturity. Even if you know the answer, show structured thinking and clear decision-making. [3] Always explain your choices, even if they’re wrong Interviews are less about the correct answer and more about your thought process. Prefer saying something logical than staying silent. For example: “I’m using a regular for loop here because it gives me better control.” [4] Practice talking while solving Most people freeze not because they lack skill  but because they don’t rehearse speaking + thinking. Get used to explaining your logic out loud while coding. Follow the rule: ABC – Always Be Chatting (not just coding). [5] Don’t sit in silence when you're stuck Silence kills momentum. If you get stuck, ask questions: “Should I print this line-by-line or return a result array?” Showing that you can stay calm under pressure matters more than getting it perfect. [6] Prefer readable, scalable solutions over clever tricks Avoid language-specific shortcuts (e.g., JS type coercion). Use modulo instead of overcomplicating divisibility logic. Interviewers prefer clean, maintainable code that works across languages. [7] Optimize code for simplicity and future changes If new conditions are added (like printing something for multiples of 7), your code should be adaptable. Use a variable-based solution instead of nested `if-else` trees. Think like an engineer, not just a coder. [8] Know your basics (like Big-O) Be ready to talk about time complexity. Even if the problem is trivial, say something like: “We’re iterating 100 times, so technically O(n), but since it's a fixed 100, it simplifies to O(1).” Technical interviews are not just a test of memory or speed. They’re also a test of clarity under pressure, structured thinking, and communication.

138

11 Comments