luke lombardi

Introducing Capsule — The Infra Framework for AI Apps Capsule is a Python framework that provides infrastructure primitives for sandboxes, auth, session management, integrations, and payments. Capsule is @supabase for AI apps. You get a powerful Python SDK to iterate fast, and one command to deploy to the cloud. Using Capsule feels like using @beam_cloud or @modal. You don’t need to set up any infrastructure. There’s no Docker, just Python. The feedback loop is super fast. You deploy to the cloud with one command. At Beam, we have a bird’s-eye view into the AI apps people are building. Over the past year, we've watched people build the same 5-7 infra primitives over and over again. Capsule provides all of those primitives using a single SDK. @llom2600 originally built Capsule to prototype our own AI apps. We started selling some of these apps for real money and realized every app needs this, not just ours. Give it a try! Curious to hear what people think. Do AI apps have their Supabase yet?

@ynnilay Looks great!

I checked my Claude history. A lot of what I type isn't questions. It's context. Backstory. Pasted emails. I'm doing the AI's job for it. So I built Contextify

are there any AI-native platform engineering / incident response tools? I want something that: - recognizes a production issue, has gated access to my k8s cluster - automatically creates a runbook with mediation steps - gets my feedback, optionally runs through the mediation steps is anyone building this?

@auglon1 @mernit Right now the payments layer is set up with stripe but is abstracted enough that most payment providers could be added pretty quickly

For the past few months, I’ve been running an AI swat team. We learn about a company, audit their workflows, and build them an AI app to automate their busywork. This sounds straightforward, but building these apps is annoying because you keep rebuilding the same things over and over: a filesystem, sandboxes per user, auth, streaming, and connections to messaging apps like Telegram and Slack So today we’re releasing all of it together in a batteries-included Python framework called Capsule It includes a sandboxed computer, multi-channel messaging, background jobs, multi-user auth, and payments – everything you need to ship a production-ready AI product super fast The faster you ship, the more you’ll learn from real customers, and the sooner you get PMF

@sockdrawermoney @muellerberndt Why not just use BAML?

I've spent every day for the last 14 months building a language for scripting LLMs because I believe we need new primitives to defend against prompt injection. Here's why: x.com/sockdrawermone…

@steipete Your code is trash tho lol. Maybe should plan more?

I never use plan mode. The main reason this was added to codex is for claude-pilled people who struggle with changing their habits. just talk with your agent.

@LangChain Why do people use this?

🚀 We just shipped a major update to LangSmith Agent Builder: • New agent chat: One always-available agent with access to all your workspace tools • Chat → Agent: Turn any conversation into a specialized agent with one click • File uploads: Attach files directly to Agent Builder • Tool registry: Add, authenticate, and manage your tools in one place Try it now: smith.langchain.com/agents?skipOnb… Learn more: blog.langchain.com/new-in-agent-b…

A Python decorator is all you need to deploy AI apps! Beam is an open-source alternative to Modal for deploying serverless AI workloads. Simply pip install beam, add a decorator, and turn any function into a serverless endpoint! 100% open-source.

@vikfand @ky__zo You can also self-host a serverless gpu runtime that provides the same UX as these providers, if you already have access to some GPUs or cloud credits: github.com/beam-cloud/bet…

@ky__zo There are many ways to run Whisper. See github.com/viktorfa/aweso…. If speed is important, use Gladia or Sieve or run your own work e.g. Modal. Replicate can be quite slow because of boot time.

I was wrong: 1) Apparently, Whisper is not one, but multiple different models from which large-v3 is the best one 2) For some reason OpenAI's API is not based on large-v3, but large-v2 (which is ambiguously called whisper-1 in the docs). There's no way to change the model 3) For whisper-1 (large-v2), I used "Prompt" (not initial prompt, but prompt) as guidance for the model of how to spell the words. I assumed it'd work if I just paste the full text of the transcript. This was the reason why sentences or individual words were skipped in the transcript. 4) I moved from OpenAI API to Replicate (large-v3) and the result is much better, although latency increased (from ~3 sec to ~10 sec for 1-minute long audio). 5) Everyone is recommending deepgram.com and I don't know it that's because it's really that good or it's just marketing.