Kyle @ KyTech

@aarondfrancis your cheque is in the mail! - 1.73GB : 5Min 1080:30 input > 140MB CPU only - 2m8s runtime @ 32vCPU (maybe overkill?) - 91% compression across 3 HLS output resolutions - Cost: $0.03..... 🤯 @modal

@aarondfrancis @modal

@KyTechInc @modal you owe me 30% of the savings, I dont make the rules

I moved my whole video rendering pipeline to @modal and couldn't be happier

@aarondfrancis @modal Napkin math checks outs. Ok back to the coding board....

@KyTechInc @modal I think like 16 or 32 vcpu

@aarondfrancis @modal woah, would not have guessed! How big is "huge" lol I think i have some testing to do. Cause TBH, the GPU is fast but its been a deps and container image hell.....

@KyTechInc @modal I went huge CPU! I did some tests and GPU wasnt fast enough to justify the costs. I assume most AI folks want GPUs, but it's nice that ffmpeg doesnt need it

@aarondfrancis @modal Now real question is what GPU did you settle on? I've been having errors with the Pro-6000 and h264_nvenc, but the modal team has a fix coming to their GPU workers that will unlock this option again. For now its the L40S

@KyTechInc @modal bingo

@ChristianLempa Dokploy > Coolify Also have you tried their CF tunnel setup? For me this was the missing piece and works so well.

When you install Dokploy, it automatically enables Docker Swarm mode. Adding a second or third node to your cluster is native and instant — it's how the tool is designed to work. Coolify treats multi-server more like managing separate remote Docker hosts. It works, but it's not the same as a real unified cluster. If you plan to scale beyond one server, this distinction matters more than any UI feature.

@AlexMontas @yazins No connection needed. Just open Claude code in the DIR where you meeting notes were saved.

@yazins Doing gods work! But I want to connect the output to my other tools (Claude code) how easy is that to do?

Just shipped the biggest update to OpenOats (prev OpenGranola: the open-source meeting copilot for macOS): Meeting Templates + AI Notes ✨ Pick a template before your meeting (1:1, Customer Discovery, Hiring, Stand-Up, Weekly) and after it ends, generate structured notes from the full transcript with one click. Notes stream in real-time as the LLM writes them. Browse past sessions, regenerate with different templates, copy to clipboard. Also fixed a subtle bug where the last ~3 seconds of audio could be lost when stopping a session. Now the app gracefully drains all buffered speech before closing. 4 new files, 1,277 lines added. All open source. github.com/yazinsai/OpenO…

@mxkaske @shadcn Both plz! 🙏

working on logs.run/docs to help you quickly integrate large data tables into your projects. it would include ~100 files, but I'm unsure at this point whether we should distribute it via the @shadcn registry or if a good agent skill would be enough. or maybe both.

@ctatedev --exclude-selector flag that works both on its own or with the current -s flag to allow for opt-in and out css selectors for complex layered scraping of pages with weird div structures/groupings.

agent-browser just hit 20k stars Really grateful for the support and feedback What is the #1 thing that would make agent-browser better for you?

@saradietschy Are you using an API key from OpenAI? They have a tiered rate limit structure based on your spend. The first two tiers have too low of a rate limit for agentic coding flows.

Ok now it's just not working at all for me. 3rd time it's happened in the last 15 min. Was so excited to get back to things after a couple weeks of focusing on YouTube. Fail. I'll try again tomorrow.

@saradietschy I would actually say its easier when using SQL/Supabase if you use MCPs or other tool calls without permission checks. A model could easily send something like a "db reset" command, wiping out all data and/or schema with one enter key....

When AI deletes an entire person's database how does that actually happen? Is it easier when db logic lives more in the code like with convex? I do DB stuff with SQL scripts or directly in supabase. Don't connect via mcp or anything -- I'm not sure how cc would destroy it but now i'm scared 😂

@JohnPhamous Any reason why the filters for features are gone? Namely the ZDR filter to only look at models/providers with zero data retention? Its the only models were allowed to use with our client base.

we improved the vercel ai gateway model table let us know any feedback you have

@adibhanna I’m gonna need more ram then lol… between the multi track recording and my local dev setup (cursor/docker/chrome/next dev) I’m running redline on my 18GB MBP

@KyTechInc yeah Continuity is capped at 1080p! with usb c you’ll get the full res! and charging

Coming soon to Lumary, Annotations with whiteboard (dots, lines, squares, and blank + colors)

@adibhanna Nice! I already thought the continuity looked good but having it on usb c would be great plus helps with charging too.

@KyTechInc it’s definitely possible, but it will take time before I can get to that! the next major feature im working on is an iPhone app that allows you to use the phone cam with its full resolution and features using usb c. (no more low quality continuity cam)

@peduarte Or replace Whisper API with Nvidias Parkeet v3 for free and honestly on a M series Mac might be the same if not faster in some cases then the API.

my bot can now transcribe youtube videos yt-dlp (audio download via dataimpulse residential proxy) → OpenAI Whisper API $5 initial top up should last ~500 vids

@peduarte RAG or projects to have a more dedicated space for larger task + not needing to load full files into the chat context. Also leveraging skills!

hacking on raycast today what's something you wish raycast could do? ai related

@amanvarshney01 @Tailscale 1 vps for now…. lol But no I through it on my Mac, and any other systems and then have 1 dashboard for all. Plus it shows docker containers too and the logs. I like Dokploy but find their metrics so so and misleading sometimes.

@KyTechInc @Tailscale I only have a single vps, is it still usefull? I can use dokploy's monitoring to see the usage 🤔

Hosted some more cool apps 😌 - wakapi: an wakatime alternative to track my coding sessions - copyparty: file server - tinyauth: using as traefik middleware to add auth (try cracking this tools.amanv.cloud)