Samrat Man Singh

I've just released Handoff, a new Elixir library for distributed function graph execution. hexdocs.pm/handoff Among other things, the library enables distributed Nx computations (anyone said distributed inference?) A huge part of this was sponsored by tvlabs.ai!

Introducing Tidewave: tidewave.ai While working on our web apps, we run code, query the database, read logs, search docs… but our AI tools are limited to reading code. Tidewave closes this gap by introducing Runtime Intelligence, for Phoenix and Rails. More coming soon, check out the website!

Testing Voice AI flows can become tedious fast But... you can just simulate conversations using another voice agent. Then, evaluate how the conversation went.

Got a small, standalone Python script working to use the new OpenAI transcription models(`gpt-4o-transcribe` and `gpt-4o-mini-transcribe`) using the Realtime WebSocket connection: samrat.me/standalone-o...

OpenAI's new TTS looks(and sounds) pretty great for the price. Also, I hope this pushes other providers to just price API usage by minute. Every other TTS provider(ElevenLabs, Cartesia, etc) currently have monthly credits pricing.

Phoenix now comes with DaisyUI. Looks like a great addition for non-designers(like me) #ElixirLang

What does talk-prep look like? Bug-fixin' mostly @samrat.me darling, you have a PR :D github.com/samrat/kokor...

Pushed my first @replicate.com model, SmolVLM2 by @hf.co (@pcuenq.hf.co et al.) Will push the 500M and 2.2B models soon as well.

Voice models are getting scary good. This one feels very human: www.sesame.com/research/cro...

I wrote up a post on how you can run AI models on Elixir that aren't yet supported by Bumblebee samrat.me/running-ml-m... #MyElixirStatus #ElixirLang

Wow, this looks super useful for running models that aren't yet supported in Bumblebee(among other things) #ElixirLang github.com/livebook-dev...

Holy smokes this is awesome, a new debugger for Phoenix LiveView. Think React devtools, but not a chrome extension. github.com/software-man... #ElixirLang

We did it! Kokoro TTS (v1.0) can now run 100% locally in your browser w/ WebGPU acceleration. Real-time text-to-speech without a server. ⚡️ Generate 10 seconds of speech in ~1 second for $0. What will you build? 🔥

Interested in how to build voice AI apps in Elixir? Here's a blog post I wrote about building a Membrane pipeline that orchestrates audio/video communication between a browser and Google Gemini #ElixirLang blog.swmansion.com/building-a-m...

Had some fun generating speech using the Elixir Kokoro implementation by @samrat.me. Made a script for typing stuff and getting the talky-talk. github.com/samrat/kokor... So a bit of ONNX, some Erlang Ports, Membrane :D #elixirlang

Reworking the Livebook for Kokoro(🔊 text-to-speech model). Now, you get a text input to input text you want read out loud(or whispered) to you. #MyElixirStatus #ElixirLang

Writing up a blog post about this. Not sure how much of it should be an introduction to Membrane. #ElixirLang

Elixir bindings for Kokoro TTS github.com/samrat/kokoro #ElixirLang

Oops, forgot to tag #ElixirLang

More voice AI exploration: Bumblebee doesn't currently support text-to-speech models, but managed to get Kokoro TTS working with Ortex Demo 👇 youtu.be/VFKX6Af9gs4

Posted on HN. Feel like sharing here as well.

I've been building a small voice AI app. So far, both OpenAI Realtime and Gemini Multimodal Live seem very unreliable with function calling. And Gemini insists that it did the function call😅

Got Gemini Multimodal Live API working with Membrane WebRTC. Here's me having a short chat with it #ElixirLang youtu.be/JxWE3pxyGqA

Boring title, but this is a delightful read on China, travel and life www.dwarkeshpatel.com/p/notes-on-c...

Trying to add interruption support to github.com/membranefram... (the API already returns events for when it detects it got interrupted, but I'm having trouble stopping playback 🤔) Currently trying to wrap my head around Membrane's manual demand mode. #ElixirLang

Current status: In Chulilla, Spain on a climbing trip.

This also makes use of MLX's lazy evaluation to provide JIT compilation of Nx functions. The README says this, but just set it as your Nx backend and compiler, and it just works!

The homepage example from Bumblebee on using BERT is now running on EMLX! We should hopefully be releasing the library in the next few days. We're just missing a really small set of Nx.Backend calbacks and some housekeeping stuff

Looking for my next job! - Based in Berlin 🇩🇪, but also open to remote. - I've been most excited about #ElixirLang lately, but also open to other tech stacks(I've worked with Clojure, Rust, JS/TS) Some stuff I've worked on recently below👇. (And I'd appreciate a repost :))

Ok, so I took this Axon example (hexdocs.pm/axon/your_fi...) for a spin on EMLX! A bit slower than Nx.BinaryBackend (runs the loop in 5s), but this is a simple model which really doesn't take advantage of a native platform. Next target is to get full Nx API coverage so we can run Bumblebee on it.

Finished reading "The MANIAC" by Benjamín Labatut. Wrote a short review: samrat.me/book-review-... #booksky

That's funny... I could swear we started this endeavor 3 days ago, not 3 weeks! Jokes aside, big shout out to @samrat.me and @uwucocoa.moe who are taking MLX coverage by stride. We're almost at a point where we can run an Axon model! Bonus: this will be plug and play with Sharding