Ella Queen

Recently I gave a talk at SF Ruby about a library we built (and recently open-sourced!) at AngelList called Zaxcel. Why'd we build it? Excel is the lingua franca of finance, but most programmatic Excel generation libraries force you to think in coordinates and formula strings. We took a different approach: model the workbook as a directed graph where data flows through named objects. Let the library handle coordinate resolution and formula composition. The result is a DSL that actually makes sense: you write formulas as Ruby expressions, reference things by name, and the type checker catches mistakes at compile time. It's revolutionized how we build financial statements. What used to take weeks now takes hours (and with AI agents - only minutes!), empowering thousands of workbook deliveries a quarter. Read how we went from nightmare refactors to 100x productivity: https://lnkd.in/es2aTttR Open-sourced repo (you can use it today!): https://lnkd.in/ecKRwcaW

91

8 Comments

Two years ago, I was designing buildings. Today, I’m helping build AI agents. At first glance, architecture and AI seem worlds apart. But the more I work in this space, the more I realize they share something fundamental. Both are about designing complex systems. Architecture designs spatial systems — how people move, how structures interact, how invisible infrastructure supports everything we see. Companies work the same way. Behind every business are operational systems: supply chains, coordination, approvals, logistics. At BackOps, we’re building AI agents to help automate those invisible workflows that keep companies running. Proud to see the team reach this Series A milestone. Check out the announcement video below 👇

13

4 Comments

✈️ I will be in SF all week for Stripe sessions. With the latest legal changes to the Apple Store's payment policies, mobile devs can now put payment processing directly into mobile applications. We can also link out to external checkout pages (Stripe checkout) Huge news for companies like Stripe and clients I support that have been begging me to put purchasing into their mobile applications for years! https://lnkd.in/ekv3Prrx https://lnkd.in/eJmVeXH5 https://lnkd.in/eGGjDFjr #iOS #Apple #Stripe

9

I’ve been loving the new work coming out of Stanford’s Digital Economy Lab and The Budget Lab at Yale -- not because it shows dramatic change, but because it’s finally asking the right questions. The most honest conclusion we have right now about AI and labor is that we simply need more (and better) measurement and a readiness to support workers even before we have definitive answers. Going to spend the next few weeks sharing more here on my read on the work. Early shifts in AI-exposed jobs, especially at the entry level, are not proof that firms are “hiring AI instead of people.” Higher exposure predicts jobs where AI could plausibly reshape how firms think about junior work: 1/ jobs where most learning can happen off the job 2/ jobs where junior and senior work/tasks look pretty similar, with speed and polish as the main differences. Those are precisely the jobs where training data already exists -- where models can be trained if labs have access to the right data. So if those jobs are changing -- in number, scope, or quality -- it could reflect many forces. It might be delayed hiring, re-sequencing of skill development, or a rethinking of what early career work even is. It does not automatically follow that senior workers have suddenly become vastly more productive thanks to AI. This (correlation != causation) matters because it has real implications for how we design policy responses. If the symptom is that early career workers are struggling, then the right interventions should be tied to those workers -- new pathways, better training models, wage protections, smoother transitions -- rather than tied narrowly to AI -- subsidizing adoption, enforcing impact assessments, or chasing productivity metrics. If we assume the wrong mechanism, we will build the wrong response. The goal isn’t to respond to "AI." It’s to respond to what’s happening to workers. Bharat Chandar, Erik Brynjolfsson, Ruyu Chen, Martha Gimbel, Molly Kinder, Joshua Kendall, Madeline Lee

243

8 Comments

Excited to share an interesting paper by our intern — a refine-and-judge pipeline that iteratively curates data for LLM fine-tuning. The approach has shown promising results on benchmarks including AlpacaEval and MT-Bench.

20

1 Comment

I recently stumbled across cocoindex-code on GitHub(https://lnkd.in/e3nJUWAq), and my reaction was… complicated. Part excitement at the “great minds think alike” moment, part quiet acknowledgment of my own bandwidth limits as a solo dev. Those who’ve followed my work know I’ve been tinkering with my own codebase-rag project last year. The core idea maps almost exactly onto what cocoindex-code does: AST-based semantic indexing, a local-first embedding design, precise retrieval to trim the context an Agent needs to consume. Seeing it executed at this level — Rust-powered engine, clean MCP integration, solid incremental indexing — my first thought was: the direction was right. My second thought was: one person really does have finite hours in a day. Over the past several months, I’ve been convinced of something that I think doesn’t get discussed enough in AI engineering circles: Agents have two hard physical constraints — context window capacity and attention bandwidth. When we ask a general-purpose Agent to reason over a 100,000-line codebase, we’re fighting both at once. My codebase-rag experiment was an attempt to work around those constraints rather than through them. cocoindex-code took the same thesis and productionized it with harder tooling. That gap got me thinking about a few broader patterns: The insight arrives before the implementation. In the LLM era, identifying the right architectural direction — like cognitive offloading — is the entry ticket. If you notice that the problem you’re independently working on is also being pursued by YC-backed teams or prolific open-source contributors, that’s a signal you’re standing on the right ground, not a reason to stop. Implementation quality determines the ceiling. I used scripts to validate the concept. The best projects are using Rust to solve production-grade performance anxiety. There’s a meaningful gap between “this works” and “this scales” — and closing that gap is where most of the craft lives. The solo developer’s real constraint isn’t ideas. It’s the constant negotiation between validating a direction and polishing an execution with finite time. Sometimes seeing a more complete implementation of your own idea isn’t discouraging — it’s a kind of release. It confirms the direction was sound, and it frees you to focus your attention on the next meaningful “how.” The Agent paradigm is shifting from instruction mode to service mode. The principle I keep coming back to: “The best context is no context.” Good AI-native applications shouldn’t make Agents heavier — they should make them lighter and more precise through well-designed external state and domain infrastructure. Respect to the open-source teams pushing this forward. And onward for the rest of us working on the architecture modernization side. #AI #AgentNative #SoftwareArchitecture #RAG #EmbeddedSystems #EngineeringMindset #OpenSource

4

Over the past year, my team and I built Pinterest's Model Context Protocol (MCP) ecosystem from the ground up: a central registry, a growing fleet of domain-specific MCP servers, and production integrations across our IDEs, internal chat surfaces, and AI agents. MCP is an open standard that gives large language models a unified way to talk to tools and data sources. At Pinterest, we used it as the substrate for AI agents that can safely automate real engineering tasks, from querying Presto data on demand to diagnosing Spark job failures and surfacing institutional knowledge. A few things I'm particularly proud of: * We designed for security from day one: two-layer auth (end-user JWTs + mesh identities), business-group-based access gating, and mandatory human-in-the-loop for sensitive actions. * We made it easy for any team to ship a new MCP server by creating a unified deployment pipeline that handles all the infrastructure, so domain experts just write business logic. * The ecosystem has scaled to 66,000+ invocations per month across 844 monthly active users, saving an estimated 7,000 engineering hours per month. This was a cross-functional effort. Thank you to everyone on the Agent Foundations, Security Engineering, and Traffic Engineering teams who made this possible, and to our engineering sponsors for their support and guidance. Full write-up on the Pinterest Engineering Blog: https://lnkd.in/e-mH26Bw

146

7 Comments

We all hate opening a full fridge and feeling like there's "nothing to eat," so I tried to fix that with a little Python. 🐍🥘 Meet Patatri's Kitchen—my personal AI assistant that tracks my pantry and plans meals based on what I actually have in stock. ✨ I used this project as a playground to dive deep into Agentic AI and Google’s Agent Development Kit (ADK). 🧠🤖 Instead of a simple chatbot, I built an intelligent "Kitchen Orchestrator" powered by Gemini 2.5 Flash. It autonomously handles the logic to: ✅ Analyze live inventory 📊 ✅ Prioritize existing ingredients to reduce waste 📉 ✅ Auto-generate shopping lists for missing items 🛒 ✅ Auto-purchase required items for cooking 🛍️ Just a fun build to experiment with orchestrating AI agents! 🛠️🚀 Check out the demo. 👇 #Python #GenerativeAI #GoogleGemini #AgenticAI #MachineLearning #GoogleADK #Coding #SideProject

22

2 Comments