I build real-time voice AI for Indian languages: speech-to-text, text-to-speech and phone agents. I like going down to the GPU kernel level to make them fast on cheap hardware. Before that I worked as a backend engineer, mostly in Go, Python and TypeScript.
- rumik-fast: I wrote my own inference engine for a 3B Hindi/English text-to-speech model.
- It runs in real time on a 4 GB laptop GPU (RTX 3050): 59 → 95–98 tokens/s, with first sound in about 0.5 s.
- Under the hood: int4 weights, hand-written CUDA int4 GEMV kernels, fused Triton decode, flash-decoding and streaming Mimi decoding.
- The int4 weights are on Hugging Face: bringagi/rumik-oss-1-int4.
- gemma-stt-hindi: a Hindi speech-to-text model built on Google's EmbeddingGemma 2 audio encoder, with a CTC head and a KenLM decoder.
- It reaches 5.4% CER on Lahaja, on par with Qwen3-ASR-1.7B.
- It was trained on free Kaggle T4 GPUs.
- The model is on Hugging Face: bringagi/gemma-stt-hindi.
- rce-go: a remote code execution service written in Go.
- http-from-scratch: HTTP/1.x implemented from scratch in Go, to learn how it works underneath.
- anymemory: a service for creating and storing persistent memories.
- terminal-todo: a small terminal todo app, in Go.
- Working on a Hindi/Hinglish end-of-turn detector, on-device Indic speech recognition, and a benchmarking harness for GPU kernels.
- Looking for roles in voice AI, inference optimization or backend. Reach me at tusharannand@gmail.com.



