Busyra Anwar

A small merchant changes their storefront theme. New layout, new hero image. Is it better? They'll never know. They don't have enough traffic to A/B test. SimGym to the rescue. Shopify x NVIDIA collaborated to build SimGym, which sends hundreds of AI-powered robot shoppers — each with a persona, a budget, and a shopping intent — to browse the store in cloud browsers. No mocked DOM. Real clicks, real scrolls, real add-to-carts. A simulated A/B test in minutes, not weeks. Under the hood: 48 NVIDIA B200 GPUs running a 120B open-source model, custom FlashInfer kernels we built with NVIDIA/CentML and contributed upstream, and 2,000 concurrent Chromium sessions on Browserbase. 94% of tokens are input. Every millisecond of inference latency is a millisecond the browser is billing. That coupling forced us to rethink how LLM serving works for agentic workloads. The result: 400,000 simulated shopping sessions per day. Cost per merchant run in single digits and falling. A merchant who's never had enough traffic for a proper test gets results in four minutes. Javier Moreno wrote up the full engineering journey — from GPT-4o API calls to self-hosted inference on Blackwell, the collaboration with NVIDIA, CentML, and the vLLM community, and the optimizations that got us here 👉 https://lnkd.in/dUesR-st Incredible work from our engineers and product leaders, including but not limited to Javier Moreno, Aaron Glazer, Alberto Castelo Becerra, Han Li, Lingyun W., Zhaoyu Z., Francis Pelland, Angelo Ocana Martins. If you're looking to build impactful AI products like SimGym, DM me or check us out @ shopify.com/careers #Shopify #AI #MachineLearning #NVIDIA #Blackwell #InfrastructureEngineering #OpenSource #vLLM #Ecommerce

14

Software engineers fear the AI that will take their job while the real danger I see should be in companies and solution owners themselves. Products that you spent years and tons of dollars for buildng selling it for whatever the number of subscriptions is can now be easily cloned and it costs nothing but the Gen AI subscription and some small resources. I am really curious of how the scene will look like after maybe Q2 of 2026. The development that business owners needs to follow “not to fail in the common nokia example” is now talking days not years like all used before!! #ArtificialIntelligence #GenerativeAI #SaaS #StartupLife #ProductManagement  #SoftwareEngineering #TechTrends #DigitalTransformation  #Innovation #FutureOfWork #Entrepreneurship  

5

2 Comments

The rate at which ClaudeCode uses mock data, hardcodes, and cheats to make a test pass or fix a bug is alarming. There’s no way you can use this tool for so-called “vibe coding” to build anything meaningful. I appreciate the progress made so far in AI-assistant coding, but these tools still have a long way to go.

4

1 Comment

If you’re a US startup hiring mobile this quarter: screen for release ownership, not Flutter trivia. Senior candidates don’t separate “engineering” from “shipping.” They own releases, hotfixes, and the messy production stuff that actually matters. One question that works: “Tell me about a release that went sideways. What broke, how did you diagnose it, and what did you change so it wouldn’t happen again?” The answer tells you if they’ve actually owned production when it broke.

17

5 Comments

Spreadsheets beat out-of-the-box software 9/10 times. Why? 1) They cost too much. (Especially in cannabis) $2500 onboarding + $1200 a month for a basic database management and visualization is absolutely insane, even with METRC API. 2) They complicate things that don’t need complicating. Changing the date of a production run requires 8 clicks in some software vs 2 in a well built sheet. 3) They lack customization. If you can’t intrinsically add / change data characteristics or dashboard attributes, (meaning without reaching out to support) the software will not grow with your company. 4) They separate data, execution, and visualization. Data sets are often hidden and reports are hard to process. The operational execution (data input) is not well integrated, and happens in a vaccum. The SOPS that drive the execution are not integrated at all The dashboards that visualize the executed data live in a different place than both the data and the operational elements. A well built spreadsheet maintains all three, live, in one place. Yes, we need software solutions for certain elements of our business, especially in a seed-to-sale tracking environment. However, people spend wayyyy too much time and energy maintaining software instead of executing their business. This is especially detrimental to sales teams that are forced to maintain complicated CRMs and outreach logs. Build your process into your data collection so it happens naturally. And don’t be afraid to do it in a spreadsheet.

33

16 Comments

Enhancing Your AI Prompts 🚀 Ever feel like you're repeating yourself in AI prompts? I've been diving into context engineering—a new trend in vibe coding—and it's a game-changer for writing production code with tools like Cursor. The secret? Manually creating files that store all those repetitive, but crucial, details. It's especially useful for tasks like migrations, where you're constantly referencing legacy code or other services. Here's my simple approach: Create a dedicated file, like legacy_code.md. Describe the context within it. Be specific about locations and key details. For example: "Legacy code for creating an order is in a separate repo outside the root: dev/legacy-order-service. In file internal/orders/service.go, you'll find the legacy implementation. For reference, the legacy DB schema is in the db_schema folder." Now, when you write your prompts, you can simply reference this file and keep your main prompt concise and focused. This saves time and makes your AI interactions more efficient. For example: "The VAT isn't calculating correctly when I submit an order. We need to fix this bug and match the legacy code behavior in @legacy_code.md." This method helps the AI get the full picture without you having to re-type the same context repeatedly. So, how do you engineer context for your daily tasks? Share your tips below! 👇

6

3 Comments

I see so many startups claiming they want to hire elite Software Engineers, but then demanding a coding task before even having an intro call. That’s not how you attract top talent.

27

17 Comments