umang

dev | ML | MLOps | LLMs | Inference

[umang @0x1umang]

Here we goo!!! lesgoo!!
@aiweekendsxyz

[ai & weekends @aiweekendsxyz]

kick off today 👀

[umang @0x1umang]

@original_ngv got his phd with overfits it seems

[umang @0x1umang]

@abhayyy_0103 hey let's connect!

[umang @0x1umang]

@ByblosDigital @X love web3xai! are you building smtn in that intersection?

[umang @0x1umang]

Testing @X algo

If you are building in
-AI
-Agents
-LLM Engineering
-ML Infrastructure
-Developer Tools
-Web3
Let’s connect. Drop what you're building currently.

[umang @0x1umang]

@rubenssoto_ai i leave testing features to myself rather than ai, maybe i’ll use small snippets of code but not the whole test prog itself from ai

[umang @0x1umang]

@querndev @X crazy build! will def try it after i finish my exam..🙃

[umang @0x1umang]

@atit_tweets @X 🔥 followed you

[umang @0x1umang]

@sutarabhiii @X heyy, followed you back!

[umang @0x1umang]

@realhisaacbrown @X up until recently I've been working on zuna - a local-first browser audiobook reader that extracts pdf/txt text and narrates it using native speech synthesis.
been in llm and inference trenches so getting some hands-on in browser side inference.

[umang @0x1umang]

@realhisaacbrown @X following you👀

[umang @0x1umang]

@edgex11lz @X sure🔥

[umang @0x1umang]

@wensitylabs @X followed you🔥

[umang @0x1umang]

piper won guys, by A LOT, will share the benchmarks 'after' my exam.

[umang @0x1umang]

taking a break from my exam prep. gonna work on my prev hackathon project now. all nighter coming up but def not for the exam.😛 https://t.co/bN5691Vkn5
what went wrong earlier?
F5-TTS delivered better custom voices, but it was too heavy and slow for browser inference.
🧵

[umang @0x1umang]

so what am I gonna do now?
VITS(Variational Inference with Adversarial Learning for e2e Text-to-Speech)and Piper TTS are apparently very good for browser side inference. will take some testing and benchmarking but I will fine tune each of em, squeeze em into smol onnx and test.

[umang @0x1umang]

I squeezed the model into a small onnx file thinking it will run, but large model files, ONNX loading, high memory use, voice switching delays, and Vercel’s static limits caused startup failures and unreliable narration, so i removed it. (zuna is inactive for now)

[umang @0x1umang]

@Shivankbuilds @X @grok dropped a follow!

[umang @0x1umang]

@akfullstack @X 🔥

[umang @0x1umang]

@boardyai @X sending dm rn!