Mohsin Mohammad

When we launched Amazon SageMaker AI in 2017, we had a clear mission: put machine learning in the hands of any developer, regardless of skill level. We wanted infrastructure engineers who were “total noobs in machine learning” to be able to achieve meaningful results in a week. To remove the roadblocks that made ML accessible only to a select few with deep expertise. Eight years later, that mission has evolved. Builders aren’t just training simple models—they’re building generative AI applications that require massive compute, complex infrastructure, and sophisticated tooling. I’ve met with customers who are training massive models, fine-tuning for specific use cases, and building applications that would have seemed like science fiction just a few years ago. But in these conversations, I hear about the same frustrations. The workarounds. The impossible choices. The time lost to what should be solved problems. A few weeks ago, we released a few new SageMaker AI capabilities that address some of the most persistent points of friction for our builders, like the SSH workaround tax. It’s the type of ongoing and incremental work that doesn’t trend on social media, but it’s the type of work that makes the lives of builders easier, one capability at a time. Read more about these updates on All Things Distributed. And as always... Now, go build! https://lnkd.in/eYwbF_q7

930

35 Comments

🚀 Introducing BharatMLStack – Meesho’s Open-Source ML Platform 🇮🇳   We’re thrilled to announce the first release of BharatMLStack, a production-grade ML platform built at Meesho and now open-sourced for the community ❤️ ! We’re launching with our Online Feature Store, battle-tested at 1M+ QPS, serving ~100 IDs per query while consistently maintaining <10ms p99 latency. Key highlights: 1. Multi-DB backend support: Plug into ScyllaDB, Dragonfly, or Redis. 2. Control Plane & Admin Ops: Built-in approvals, role-based access, and audit trails. 3. Feature Cataloging UI: Manage, discover, and evolve features collaboratively. 4. Optimized for economics: Every component is designed with cost-efficiency and scale in mind.  And this is just the beginning. A whole armory of components—from stream-based feature computation to embedding search and model serving—is on its way. We're building BharatMLStack to make high-scale ML accessible, economical, and open. Dive in: https://lnkd.in/g5itp4Sk Docs: https://lnkd.in/gEimvDMx Blogs: https://lnkd.in/gZwhgrJY Quick-start: https://lnkd.in/g5zXXr27 #opensource #mlplatform #featurestore #mlops #bharatmlstack #meesho #scylladb #redis #dragonfly #machinelearning #costefficientML

385

16 Comments

The biggest difference between Google and Waymo, almost one year in: Waymo has largely adopted the Amazonian practice of presenting information in documents, not slides. Meetings start with a few minutes of pre-read / frantic commenting on a Google Doc, followed by a live burn-down of the comments that require discussion. It's efficient and information-dense. I love it.

2,183

63 Comments

Amazon Web Services (AWS) 𝗕𝗲𝗱𝗿𝗼𝗰𝗸 𝗔𝗴𝗲𝗻𝘁𝗖𝗼𝗿𝗲 𝗠𝗲𝗺𝗼𝗿𝘆 — 𝗠𝗲𝗺𝗼𝗿𝘆 𝗳𝗼𝗿 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗔𝗜 🚀 One of the biggest challenges in building AI agents has always been 𝗺𝗲𝗺𝗼𝗿𝘆. LLM-based agents are naturally 𝘀𝘁𝗮𝘁𝗲𝗹𝗲𝘀𝘀 — they forget past interactions, repeat themselves, and force developers to hack together logs, vector DBs, and retrieval systems just to simulate memory. With 𝗔𝗴𝗲𝗻𝘁𝗖𝗼𝗿𝗲 𝗠𝗲𝗺𝗼𝗿𝘆, AWS introduces a 𝗻𝗮𝘁𝗶𝘃𝗲, 𝗳𝘂𝗹𝗹𝘆 𝗺𝗮𝗻𝗮𝗴𝗲𝗱 𝗺𝗲𝗺𝗼𝗿𝘆 layer inside Bedrock. It blends short-term “working memory” with long-term structured insights, enabling agents to become context-aware and evolve across sessions. 🔍 𝗛𝗼𝘄 𝗶𝘁 𝘄𝗼𝗿𝗸𝘀 1️⃣ 𝗠𝗲𝗺𝗼𝗿𝘆 𝗥𝗲𝘀𝗼𝘂𝗿𝗰𝗲 A secure, logical container where both raw events and processed long-term memories live. You can configure retention (e.g., TTL up to 365 days) and encryption with AWS-managed or customer-managed KMS. 2️⃣ 𝗦𝗵𝗼𝗿𝘁-𝗧𝗲𝗿𝗺 𝗠𝗲𝗺𝗼𝗿𝘆 (Events) Every interaction — user inputs, agent replies, tool calls — is captured as an event. These are structured with identifiers like memoryId, actorId, and sessionId, ensuring scoped, relevant recall. This is the agent’s “working memory” for coherence in live conversations. 3️⃣ 𝗟𝗼𝗻𝗴-𝗧𝗲𝗿𝗺 𝗠𝗲𝗺𝗼𝗿𝘆 (Insights) Events are asynchronously processed into structured knowledge: preferences, semantic facts, or session summaries. These durable memories, organized in namespaces, provide the agent with a persistent sense of history. 4️⃣ 𝗠𝗲𝗺𝗼𝗿𝘆 𝗦𝘁𝗿𝗮𝘁𝗲𝗴𝗶𝗲𝘀 & 𝗡𝗮𝗺𝗲𝘀𝗽𝗮𝗰𝗲𝘀 Built-in strategies include semantic memory, summarization, and user preference capture. Developers can also define custom strategies. Namespaces act like folder paths (e.g., /customer/{actorId}/preferences) to group, isolate, and control memory. 5️⃣ 𝗔𝗱𝘃𝗮𝗻𝗰𝗲𝗱 𝗖𝗮𝗽𝗮𝗯𝗶𝗹𝗶𝘁𝗶𝗲𝘀  • 𝗕𝗿𝗮𝗻𝗰𝗵𝗶𝗻𝗴 ➝ Explore alternate conversation paths or “what-if” scenarios without losing the original thread.  • 𝗖𝗵𝗲𝗰𝗸𝗽𝗼𝗶𝗻𝘁𝗶𝗻𝗴 ➝ Mark states in workflows or sessions so agents can resume later seamlessly. 🌟 𝗪𝗵𝘆 𝘁𝗵𝗶𝘀 𝗺𝗮𝘁𝘁𝗲𝗿𝘀 𝗔𝗴𝗲𝗻𝘁𝗖𝗼𝗿𝗲 𝗠𝗲𝗺𝗼𝗿𝘆 turns agents from stateless to stateful. Instead of disconnected sessions, we now get continuous, evolving relationships where AI can:  • Recall preferences & history  • Stay coherent over long conversations  • Resume tasks across sessions  • Explore alternates while preserving context This isn’t just storage — it’s i𝗻𝗳𝗿𝗮𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲 𝗳𝗼𝗿 𝗶𝗻𝘁𝗲𝗹𝗹𝗶𝗴𝗲𝗻𝗰𝗲. For copilots, assistants, or domain-specific advisors, memory is no longer optional — it’s mission-critical. With 𝗔𝗴𝗲𝗻𝘁𝗖𝗼𝗿𝗲 𝗠𝗲𝗺𝗼𝗿𝘆, developers don’t need to reinvent complex pipelines. They can focus on what matters most: building agents that feel trustworthy, reliable, and human-like. https://lnkd.in/gMvaJzf3 #AgenticAI #AI #AgenticMemory #AWS

70

5 Comments

Last week, OpenAI committed $300B in cloud compute orders to Oracle. Oracle was up 30%, adding over $200B to its market cap. How will Oracle get this compute? By buying chips from Nvidia. Today, Nvidia has announced that they will invest up to $100B in OpenAI to increase their compute capacity. Stock is up 4% already. So OpenAI is using some of Nvidia's money to pay Oracle, who will then pay Nvidia to buy GPUs, so that Nvidia will then invest some more money into...OpenAI. If it looks like a bubble and acts like a bubble,...

218

4 Comments

I didn’t stay and grow at Amazon for 5 years just because I could write good code. Plenty of smart engineers can do that every single day. Writing clean code was just the entry ticket. What truly set me apart was taking ownership, even when it wasn’t asked. If I was assigned a bug to fix, I could’ve stopped after the code review. But I would follow it to production, sit with support teams, talk to impacted customers, and make sure the bug was truly gone, and people felt the difference. If I was asked to build a new API, I could’ve stopped after passing the unit tests. But I’d stress-test it with real traffic, set up alerting, write docs, and hop on calls with downstream teams to make sure their integration was smooth. If I noticed on-call alerts going off every night, I could’ve shrugged and marked them as “infra issues.” But I dug into root causes, ran post-mortems, and sometimes pulled late nights re-architecting fragile parts, so the next person on call wouldn’t have to lose sleep. If I wrapped up my sprint tasks ahead of schedule, I could’ve clocked out early. But I’d sit with teammates struggling with blockers, offer code reviews, and even pick up “thankless” chores like improving internal tools or onboarding docs. Ownership isn’t about heroics or working 14-hour days. It’s about caring enough to see things through, to fix the root, not just the symptoms. To ask, “What else could go wrong?” and, “How can I make this easier for the next person?” That’s how you get noticed. That’s how you build trust. That’s how you grow. Don’t just be the engineer who writes code. Be the one who owns the outcome, end to end. That’s what really sets you apart.

297

10 Comments

Check out this great interview with Xintong Song from Alibaba Cloud on Flink Agents, a new sub-project we collaborated on to bring AI agents into Apache Flink. AI demos are easy. Production is hard. Real production AI needs to be: •  Distributed •  Scalable •  Resilient •  Event-driven Flink already solves these problems. It’s made for real-time systems that process financial transactions, clickstreams, IoT events, and more. Most agent frameworks? They either try to rebuild these capabilities from scratch or you have to bolt multiple systems together. Flink Agents flips this model. Agents are brought to the data and signals that matter, embedded in the stream, not the other way around. Instead of agents waiting for a user to ask something, they react to system events, the heartbeat of your business. Imagine a livestream with thousands of comments per second. Today, a backup team might try to monitor them. With Flink Agents, you can deploy agents that summarize and react in real-time, no human supervision required. It’s a natural fit. Agents need everything Flink already does best. Watch the talk here to see it in action: https://lnkd.in/gnrxKrsi #Flink #AI #StreamingAgents #EventDrivenAI #OpenSource #ApacheFlink

28

Google Announces First AI Hub in India 🇮🇳, Bringing Company’s Full AI Stack and Consumer Services to Country. This investment of approximately $15 billion (USD) over five years (2026-2030) is Google’s largest investment in India to date, and is aligned to the Indian Government’s Viksit Bharat 2047 vision to accelerate the expansion of AI-driven services. The Google AI hub initiative includes: 🤖 A Gigawatt-Scale Data Center Campus purpose-built to handle the demands of next-generation AI, powering services like Search, Gemini, Workspace, and YouTube for millions of users across India and the globe. 🛜 A new International Subsea Gateway on India's East Coast, which will include multiple new subsea cables. This will strategically deliver high-speed, resilient connectivity and bolster the entire nation's digital backbone. ⚡ Clean Energy Infrastructure to ensure we grow responsibly and advance India’s clean energy goals. Check out: https://lnkd.in/gDQjNH7C

63

Excited to share that Shweta Garg and I will be speaking at Uber! Our talk, "From Code to Impact: AI Teammates for Developer Productivity and Reliability," dives into how collaborative coding agents can accelerate developer workflows, reduce toil, and improve performance, reliability, and code quality at scale. We will explore the building blocks of effective AI teammates (including planning, memory, tool use, and multi-agent collaboration) and show how these capabilities can transform GenAI-powered programming systems for large-scale engineering teams. #Uber #GenerativeAI #AIAgents #DeveloperProductivity #CodeQuality

36