Engineering

AI raised the floor. Engineering excellence raises the ceiling. It's so riveting to see new LLM models get published and the step changes that are happening. AI has made it dramatically easier to produce code. It has simultaneously made it much harder to hide weak engineering fundamentals. AI is raising the floor, meaning more people can generate software and prototypes quickly. But engineering excellence raises the ceiling: determining whether that code becomes a reliable, scalable system that actually creates enterprise value. AI is exposing something many organizations have quietly carried for years: technical debt, fragile architectures, and disconnected data foundations. When systems aren't built well, AI doesn't fix that. It simply reveals it faster. 💡  𝗜 𝗮𝗺 𝗮 𝘀𝘁𝗿𝗼𝗻𝗴 𝗯𝗲𝗹𝗶𝗲𝘃𝗲𝗿 𝘁𝗵𝗮𝘁 𝘁𝗼 𝗺𝗮𝘅𝗶𝗺𝗶𝘇𝗲 𝗔𝗜 𝘃𝗮𝗹𝘂𝗲, 𝘄𝗲 𝗻𝗲𝗲𝗱 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 𝗲𝘅𝗰𝗲𝗹𝗹𝗲𝗻𝗰𝗲. So what does engineering excellence look like right now? I think about it as four pillars: ▸ 𝗔𝗜-𝗥𝗲𝗮𝗱𝘆 𝗔𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲: AI doesn't work well on top of poor architecture. Modernizing legacy code without addressing underlying structure just produces the wrong architecture faster. ▸ 𝗛𝗶𝗴𝗵-𝗤𝘂𝗮𝗹𝗶𝘁𝘆 𝗗𝗮𝘁𝗮 𝗙𝗼𝘂𝗻𝗱𝗮𝘁𝗶𝗼𝗻𝘀: AI is only as intelligent as the data it reasons over. You can't shortcut this layer and even a strong foundation must continuously evolve. ▸ 𝗦𝗲𝗰𝘂𝗿𝗲 𝗮𝗻𝗱 𝗢𝗯𝘀𝗲𝗿𝘃𝗮𝗯𝗹𝗲 𝗦𝘆𝘀𝘁𝗲𝗺𝘀: As AI agents become more autonomous, seeing what's happening and why becomes non-negotiable. Governance isn't just policy it's instrumentation and operationalization, as many of you noted in my last post. ▸ 𝗗𝗶𝘀𝗰𝗶𝗽𝗹𝗶𝗻𝗲𝗱 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 𝗣𝗿𝗮𝗰𝘁𝗶𝗰𝗲𝘀: Spec discipline, test rigor, strong code review, clear ownership are not legacy practices to abandon, but more important than ever. AI rewards good fundamentals and makes the consequences of weak ones more visible, faster. There's a real shift in how engineers spend their time. Less writing foundational code. More orchestrating systems: designing architecture, shaping how AI agents interact, validating outputs with genuine judgment. I see our senior engineers flying because their systems thinking depth makes AI a true force multiplier. Earlier-career engineers are learning, but need more deliberate mentorship than ever. When AI can simulate senior output, the risk is gaining confidence without gaining understanding. The best thing leaders can do: create conditions where engineers are proud of how they build, not just what they ship. The time savings alone aren't the win. For us, we are investing in deeper architecture work, stronger data foundations, the next generation of agentic capabilities and I believe that's the winning combo. 𝗗𝗼 𝘆𝗼𝘂 𝗮𝗴𝗿𝗲𝗲 𝘁𝗵𝗮𝘁 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 𝗲𝘅𝗰𝗲𝗹𝗹𝗲𝗻𝗰𝗲 𝗶𝘀 𝗺𝗼𝗿𝗲 𝗶𝗺𝗽𝗼𝗿𝘁𝗮𝗻𝘁 𝘁𝗵𝗮𝗻 𝗲𝘃𝗲𝗿?

Scaling from 50 to 100 employees almost killed our company. Until we discovered a simple org structure that unlocked $100M+ in annual revenue. In my 10+ years of experience as a founder, one of the biggest challenges I faced in scaling was bridging the organizational gap between startup and enterprise. We hit that wall at around 100~ employees. What worked beautifully with a small team suddenly became our biggest obstacle to growth. The problem was our functional org structure: Engineers reporting to engineering, product to product, business to business. This created a complex dependency web: • Planning took weeks • No clear ownership  • Business threw Jira tickets over the fence and prayed for them to get completed • Engineers didn’t understand priorities and worked on problems that didn’t align with customer needs That was when I studied Amazon's Single-Threaded Owner (STO) model, in which dedicated GMs run independent business units with their own cross-functional teams and manage P&L It looked great for Amazon's scale but felt impossible for growing companies like ours. These 2 critical barriers made it impractical for our scale: 1. Engineering Squad Requirements: True STO demands complete engineering teams (including managers) reporting to a single owner. At our size, we couldn't justify full engineering squads for each business unit. To make it work, we would have to quadruple our engineering headcount. 2. P&L Owner Complexity: STO leaders need unicorn-level skills: deep business acumen and P&L management experience. Not only are these leaders rare and expensive, but requiring all these skills in one person would have limited our talent pool and slowed our ability to launch new initiatives. What we needed was a model that captured STO's focus and accountability but worked for our size and growth needs. That's when we created Mission-Aligned Teams (MATs), a hybrid model that changed our execution (for good) Key principles: • Each team owns a specific mission (e.g., improving customer service, optimizing payment flow) • Teams are cross-functional and self-sufficient,  • Leaders can be anyone (engineer, PM, marketer) who's good at execution • People still report functionally for career development • Leaders focus on execution, not people management The results exceeded our highest expectations: New MAT leads launched new products, each generating $5-10M in revenue within a year with under 10 person teams. Planning became streamlined. Ownership became clear. But it's NOT for everyone (like STO wasn’t for us) If you're under 50 people, the overhead probably isn't worth it. If you're Amazon-scale, pure STO might be better. MAT works best in the messy middle: when you're too big for everyone to be in one room but too small for a full enterprise structure. image courtesy of Manu Cornet ------ If you liked this, follow me Henry Shi as I share insights from my journey of building and scaling a  $1B/year business.

Exciting updates on Project GR00T! We discover a systematic way to scale up robot data, tackling the most painful pain point in robotics. The idea is simple: human collects demonstration on a real robot, and we multiply that data 1000x or more in simulation. Let’s break it down: 1. We use Apple Vision Pro (yes!!) to give the human operator first person control of the humanoid. Vision Pro parses human hand pose and retargets the motion to the robot hand, all in real time. From the human’s point of view, they are immersed in another body like the Avatar. Teleoperation is slow and time-consuming, but we can afford to collect a small amount of data.  2. We use RoboCasa, a generative simulation framework, to multiply the demonstration data by varying the visual appearance and layout of the environment. In Jensen’s keynote video below, the humanoid is now placing the cup in hundreds of kitchens with a huge diversity of textures, furniture, and object placement. We only have 1 physical kitchen at the GEAR Lab in NVIDIA HQ, but we can conjure up infinite ones in simulation. 3. Finally, we apply MimicGen, a technique to multiply the above data even more by varying the *motion* of the robot. MimicGen generates vast number of new action trajectories based on the original human data, and filters out failed ones (e.g. those that drop the cup) to form a much larger dataset. To sum up, given 1 human trajectory with Vision Pro  -> RoboCasa produces N (varying visuals)  -> MimicGen further augments to NxM (varying motions). This is the way to trade compute for expensive human data by GPU-accelerated simulation. A while ago, I mentioned that teleoperation is fundamentally not scalable, because we are always limited by 24 hrs/robot/day in the world of atoms. Our new GR00T synthetic data pipeline breaks this barrier in the world of bits. Scaling has been so much fun for LLMs, and it's finally our turn to have fun in robotics! We are creating tools to enable everyone in the ecosystem to scale up with us: - RoboCasa: our generative simulation framework (Yuke Zhu). It's fully open-source! Here you go: http://robocasa.ai - MimicGen: our generative action framework (Ajay Mandlekar). The code is open-source for robot arms, but we will have another version for humanoid and 5-finger hands: https://lnkd.in/gsRArQXy - We are building a state-of-the-art Apple Vision Pro -> humanoid robot "Avatar" stack. Xiaolong Wang group’s open-source libraries laid the foundation: https://lnkd.in/gUYye7yt - Watch Jensen's keynote yesterday. He cannot hide his excitement about Project GR00T and robot foundation models! https://lnkd.in/g3hZteCG Finally, GEAR lab is hiring! We want the best roboticists in the world to join us on this moon-landing mission to solve physical AGI: https://lnkd.in/gTancpNK

Last week, I described four design patterns for AI agentic workflows that I believe will drive significant progress: Reflection, Tool use, Planning and Multi-agent collaboration. Instead of having an LLM generate its final output directly, an agentic workflow prompts the LLM multiple times, giving it opportunities to build step by step to higher-quality output. Here, I'd like to discuss Reflection. It's relatively quick to implement, and I've seen it lead to surprising performance gains. You may have had the experience of prompting ChatGPT/Claude/Gemini, receiving unsatisfactory output, delivering critical feedback to help the LLM improve its response, and then getting a better response. What if you automate the step of delivering critical feedback, so the model automatically criticizes its own output and improves its response? This is the crux of Reflection. Take the task of asking an LLM to write code. We can prompt it to generate the desired code directly to carry out some task X. Then, we can prompt it to reflect on its own output, perhaps as follows: Here’s code intended for task X: [previously generated code] Check the code carefully for correctness, style, and efficiency, and give constructive criticism for how to improve it. Sometimes this causes the LLM to spot problems and come up with constructive suggestions. Next, we can prompt the LLM with context including (i) the previously generated code and (ii) the constructive feedback, and ask it to use the feedback to rewrite the code. This can lead to a better response. Repeating the criticism/rewrite process might yield further improvements. This self-reflection process allows the LLM to spot gaps and improve its output on a variety of tasks including producing code, writing text, and answering questions. And we can go beyond self-reflection by giving the LLM tools that help evaluate its output; for example, running its code through a few unit tests to check whether it generates correct results on test cases or searching the web to double-check text output. Then it can reflect on any errors it found and come up with ideas for improvement. Further, we can implement Reflection using a multi-agent framework. I've found it convenient to create two agents, one prompted to generate good outputs and the other prompted to give constructive criticism of the first agent's output. The resulting discussion between the two agents leads to improved responses. Reflection is a relatively basic type of agentic workflow, but I've been delighted by how much it improved my applications’ results. If you’re interested in learning more about reflection, I recommend: - Self-Refine: Iterative Refinement with Self-Feedback, by Madaan et al. (2023) - Reflexion: Language Agents with Verbal Reinforcement Learning, by Shinn et al. (2023) - CRITIC: Large Language Models Can Self-Correct with Tool-Interactive Critiquing, by Gou et al. (2024) [Original text: https://lnkd.in/g4bTuWtU ]

When working with multiple LLM providers, managing prompts, and handling complex data flows — structure isn't a luxury, it's a necessity. A well-organized architecture enables: → Collaboration between ML engineers and developers → Rapid experimentation with reproducibility → Consistent error handling, rate limiting, and logging → Clear separation of configuration (YAML) and logic (code) 𝗞𝗲𝘆 𝗖𝗼𝗺𝗽𝗼𝗻𝗲𝗻𝘁𝘀 𝗧𝗵𝗮𝘁 𝗗𝗿𝗶𝘃𝗲 𝗦𝘂𝗰𝗰𝗲𝘀𝘀 It’s not just about folder layout — it’s how components interact and scale together: → Centralized configuration using YAML files → A dedicated prompt engineering module with templates and few-shot examples → Properly sandboxed model clients with standardized interfaces → Utilities for caching, observability, and structured logging → Modular handlers for managing API calls and workflows This setup can save teams countless hours in debugging, onboarding, and scaling real-world GenAI systems — whether you're building RAG pipelines, fine-tuning models, or developing agent-based architectures. → What’s your go-to project structure when working with LLMs or Generative AI systems? Let’s share ideas and learn from each other.

Should you try Google’s famous “20% time” experiment to encourage innovation? We tried this at Duolingo years ago. It didn’t work. It wasn’t enough time for people to start meaningful projects, and very few people took advantage of it because the framework was pretty vague. I knew there had to be other ways to drive innovation at the company. So, here are 3 other initiatives we’ve tried, what we’ve learned from each, and what we're going to try next. 💡 Innovation Awards: Annual recognition for those who move the needle with boundary-pushing projects. The upside: These awards make our commitment to innovation clear, and offer a well-deserved incentive to those who have done remarkable work. The downside: It’s given to individuals, but we want to incentivize team work. What’s more, it’s not necessarily a framework for coming up with the next big thing. 💻 Hackathon: This is a good framework, and lots of companies do it. Everyone (not just engineers) can take two days to collaborate on and present anything that excites them, as long as it advances our mission or addresses a key business need. The upside: Some of our biggest features grew out of hackathon projects, from the Duolingo English Test (born at our first hackathon in 2013) to our avatar builder. The downside: Other than the time/resource constraint, projects rarely align with our current priorities. The ones that take off hit the elusive combo of right time + a problem that no other team could tackle. 💥 Special Projects: Knowing that ideal equation, we started a new program for fostering innovation, playfully dubbed DARPA (Duolingo Advanced Research Project Agency). The idea: anyone can pitch an idea at any time. If they get consensus on it and if it’s not in the purview of another team, a cross-functional group is formed to bring the project to fruition. The most creative work tends to happen when a problem is not in the clear purview of a particular team; this program creates a path for bringing these kinds of interdisciplinary ideas to life. Our Duo and Lily mascot suits (featured often on our social accounts) came from this, as did our Duo plushie and the merch store. (And if this photo doesn't show why we needed to innovate for new suits, I don't know what will!) The biggest challenge: figuring out how to transition ownership of a successful project after the strike team’s work is done. 👀 What’s next? We’re working on a program that proactively identifies big picture, unassigned problems that we haven’t figured out yet and then incentivizes people to create proposals for solving them. How that will work is still to be determined, but we know there is a lot of fertile ground for it to take root. How does your company create an environment of creativity that encourages true innovation? I'm interested to hear what's worked for you, so please feel free to share in the comments! #duolingo #innovation #hackathon #creativity #bigideas

The old approach of sending resumes and hoping for the best isn't working anymore. Thousands of talented engineers are competing for fewer positions. In this market, being skilled isn't enough. You need to be visible. The engineers who are landing roles fast aren't necessarily the most qualified. They're the ones who know how to promote themselves and stand out from the crowd. That's why I created this 5-𝘀𝘁𝗲𝗽 𝗮𝘁𝘁𝗿𝗮𝗰𝘁𝗶𝗼𝗻 𝘀𝘆𝘀𝘁𝗲𝗺 𝘁𝗼 𝗵𝗲𝗹𝗽 𝘆𝗼𝘂 𝗿𝗶𝘀𝗲 𝗮𝗯𝗼𝘃𝗲 𝘁𝗵𝗲 𝗻𝗼𝗶𝘀𝗲: 📍 Step 1: Optimize Your LinkedIn Profile ↳ Your headline should immediately showcase your specific expertise. ↳ Quantify your achievements. ↳ Make yourself discoverable when recruiters search. 📍 Step 2: Build a Killer GitHub Portfolio ↳ Create 3-4 production-grade projects with detailed READMEs. ↳ Show your thinking process. ↳ Prove your skills instead of just listing them. 📍 Step 3: Write Technical Content Document what you learn. ↳ Share project walkthroughs. ↳ Write about common mistakes. 📍 Step 4: Share Strategically Post your insights with context. ↳ Explain why topics matter. ↳ Document your learning journey consistently. 📍 Step 5: Grow Your Network ↳ Connect with recruiters proactively. ↳ Engage meaningfully with posts daily. ↳ Build relationships before you need them. The result: Instead of competing with hundreds of identical resumes, you become the engineer they already know and want to hire. This system works because it positions you as a known solution, not an unknown candidate. 📌 Want the complete breakdown with actionable tips? Download the full guide here: https://bit.ly/4mZk17A I really hope this is useful. Share this with someone in your network who could benefit from these strategies. 💬 What's the biggest challenge you're facing in this competitive market?

It’s easy as a PM to only focus on the upside. But you'll notice: more experienced PMs actually spend more time on the downside. The reason is simple: the more time you’ve spent in Product Management, the more times you’ve been burned. The team releases “the” feature that was supposed to change everything for the product - and everything remains the same. When you reach this stage, product management becomes less about figuring out what new feature could deliver great value, and more about de-risking the choices you have made to deliver the needed impact. -- To do this systematically, I recommend considering Marty Cagan's classical 4 Risks. 𝟭. 𝗩𝗮𝗹𝘂𝗲 𝗥𝗶𝘀𝗸: 𝗧𝗵𝗲 𝗦𝗼𝘂𝗹 𝗼𝗳 𝘁𝗵𝗲 𝗣𝗿𝗼𝗱𝘂𝗰𝘁 Remember Juicero? They built a $400 Wi-Fi-enabled juicer, only to discover that their value proposition wasn’t compelling. Customers could just as easily squeeze the juice packs with their hands. A hard lesson in value risk. Value Risk asks whether customers care enough to open their wallets or devote their time. It’s the soul of your product. If you can’t be match how much they value their money or time, you’re toast. 𝟮. 𝗨𝘀𝗮𝗯𝗶𝗹𝗶𝘁𝘆 𝗥𝗶𝘀𝗸: 𝗧𝗵𝗲 𝗨𝘀𝗲𝗿’𝘀 𝗟𝗲𝗻𝘀 Usability Risk isn't about if customers find value; it's about whether they can even get to that value. Can they navigate your product without wanting to throw their device out the window? Google Glass failed not because of value but usability. People didn’t want to wear something perceived as geeky, or that invaded privacy. Google Glass was a usability nightmare that never got its day in the sun. 𝟯. 𝗙𝗲𝗮𝘀𝗶𝗯𝗶𝗹𝗶𝘁𝘆 𝗥𝗶𝘀𝗸: 𝗧𝗵𝗲 𝗔𝗿𝘁 𝗼𝗳 𝘁𝗵𝗲 𝗣𝗼𝘀𝘀𝗶𝗯𝗹𝗲 Feasibility Risk takes a different angle. It's not about the market or the user; it's about you. Can you and your team actually build what you’ve dreamed up? Theranos promised the moon but couldn't deliver. It claimed its technology could run extensive tests with a single drop of blood. The reality? It was scientifically impossible with their tech. They ignored feasibility risk and paid the price. 𝟰. 𝗩𝗶𝗮𝗯𝗶𝗹𝗶𝘁𝘆 𝗥𝗶𝘀𝗸: 𝗧𝗵𝗲 𝗠𝘂𝗹𝘁𝗶-𝗗𝗶𝗺𝗲𝗻𝘀𝗶𝗼𝗻𝗮𝗹 𝗖𝗵𝗲𝘀𝘀 𝗚𝗮𝗺𝗲 (Business) Viability Risk is the "grandmaster" of risks. It asks: Does this product make sense within the broader context of your business? Take Kodak for example. They actually invented the digital camera but failed to adapt their business model to this disruptive technology. They held back due to fear it would cannibalize their film business. -- This systematic approach is the best way I have found to help de-risk big launches. How do you like to de-risk?

🪂 How To Make Your Design System AI-Ready (https://lnkd.in/dtnpy7CM), a practical guide on how to reduce drifts, minimize mistakes, maintain context and improve the quality of AI-generated prototypes — with structured spec files, automated auditing and token layers. Put together by Hardik Pandya from Atlassian. --- 🔹 1. Design Decisions Are Infrastructure AI-generated prototypes often don't deliver consistently decent results because of tiny inconsistencies scattered all across a design system. Often it's decisions made but not documented, hard-coded values never cleaned up, or relying too much on AI making sense of mock-ups or design flows on its own. Unsurprisingly, better AI prototypes come from better data — but also from better human guidance. We shouldn’t assume that AI knows how to choose the right component, and how to design with accessibility in mind. It needs priorities, a clear path on how we make decisions, design principles, examples, do's and don'ts. In fact, we should treat design decisions as infrastructure. That means that every time we make a decision — not just a design decision, but even decision on how actually prioritize our work and how we make decisions around here — it must find a path into the spec file that is then consumed by AI. --- 🔶 2. Three Layers: Spec Files + Token Layer + Audit To ensure quality, we establish design principles, guidelines, rules in a form of “spec files”). It's structured Markdown files that include spacing rules, color choices, component usage guidelines, priorities etc. AI is going to read and reuse that spec file every time it's going to generate a prototype. Because the spec files are text files, it's much more cost-effective, but also much more accurate just because we don't rely on AI recognizing or decoding patterns from mock-ups, but gets specific guidelines instead. In fact, extending code is often a more effective way than generating code from mock-ups. Token layer lists and keeps updated all tokens used throughout the design system. AI always chooses from a closed set of named variables instead of inventing plausible values ad-hoc. An audit script catches what AI gets wrong. It scans the prototype and flags every hard-coded value and flags it if necessary. It can be a regular software doing that, with AI waiting for its feedback to come back. Finally, when a design system ships updates, a sync routine flags which spec files need updating. The goal is to make sure that AI always reads up-to-date, current specs, not the ones written against an outdated version. --- 🔺 3. Examples of AI-Ready Design Systems ⌾ Atlassian: https://lnkd.in/dVsGc3Cp ⌾ Carbon: https://lnkd.in/d4zq4WWb ⌾ CMS Design System: https://lnkd.in/dHHzV3en ⌾ Nordhealth: https://lnkd.in/d8C4j2ZA Yet again, AI can’t magically resolve technical debt or design debt — it needs guidance, decisions, priorities and principles.

April 6th: A bright spring day in Germany, one that perfectly illustrates the need for battery storage systems. Like so many other sunny days, PV generation in Germany covered a large portion of the electricity demand for several hours in the middle of the day, thanks to the cloudless sky and millions of solar modules. But there is a darker side to the sunshine. Large amounts of daytime solar can overload the grid and cause severe electricity price fluctuations: on April 6th, intraday electricity prices dropped to -200€/MWh at their lowest point. In cases where more electricity is generated from solar energy than the grid can handle, grid operators regularly require solar installations to curtail their production. This means that energy that could otherwise be made available to consumers cannot be used. And when the sun goes down, most of the demand must quickly be met with flexible sources. This adds an extra layer of complexity: deciding which conventional power plants can be shut down during the day and switched on again in the evening is a careful balancing act. This is precisely the situation where battery energy storage systems (BESS) can bridge the gap, with several advantages: - By storing part of the solar energy at peak generation times and dispatching it later, BESS can help shift the curve to more closely align with evening demand. - Better management of volatile generation from renewables also helps keep prices stable. - Provided they are close to the overproducing solar systems, BESS contribute to grid stability by helping balance supply and demand. Of course, there is no one-size-fits-all technology. A secure and flexible energy system needs a diverse mix. But batteries are playing an increasing role, especially as they become more and more affordable. We at RWE are harnessing the benefits: we have 1.2 GW of installed BESS capacity worldwide, of which nine systems totalling 364 MW of capacity operate in Germany alone. We’re scaling fast, with new large-scale projects recently commissioned in Germany and the Netherlands. And we have just decided to build a BESS facility in Hamm with an installed capacity of 600 megawatts. So, let’s continue to make the most of those sunny days — by creating the right framework conditions to build up affordable and flexible support.