
Building real-time voice agents @VaaniHQ • LLM infra • Agent systems
Bengaluru, India
Joined September 2020
- Tweets219
- Following743
- Followers72
- Likes476
At Bengaluru Tech Summit this Thursday?
We’re running a hands-on Voice AI workshop. In three hours, you’ll build your own voice agent on VAANI. No code needed.
19 Nov · 10 AM–1 PM · BIEC
BTS 2026 pass required.
Register: bharat1.ai/workshop
Even at 19 ms, Gemma misses the 10 ms budget, so the small model is still the pick for the tightest path. Full story with stickman drawings, t-SNE maps and every dead end: mutedmiles.space/post/fallin…
27
I also wrote a matrix multiply in C++ six ways: 2122 ms down to 22 ms. Swapping two loops alone gave 3.5x. One clever idea (cache tiles) did nothing.
1
13
8-bit maths plus a shorter prompt prefix got me to about 19 ms. 4-bit was slower. OpenVINO with 8-bit was 8x slower. Measure, don’t assume.
1
16
My own report had said OpenVINO was 6.4x faster. Most of that was this same bug. I went back and corrected it. Profile first, then change something.
1
14
The real culprit: the model file stored weights as 8-bit, but ONNX Runtime converted ALL of them back to 32-bit on every call. One offline script fixed it: 147 to 171 ms down to 34 ms, identical embeddings.
1
9
Rust vs Python? A tie. More than 99.8% of the time is inside the model. The language was never the problem.
1
14
EmbeddingGemma picks the right intent 96.5% of the time on my 202 test sentences in 6 languages. The small model gets about 91%. But Gemma took 150 to 170 ms per sentence. A voice budget is about 10 ms.
1
21
I only wanted to turn a sentence into a vector for a voice assistant. Then I fell through nine layers: tokens, model, runtimes, kernels, CPU lanes.
What I found, in order:
1
2
96
Somil retweeted
Vaani will be at Bengaluru Tech Summit 2026! 🎉
Come hear it for yourself: voice agents making and taking real phone calls in the languages India speaks.
- 17 to 19 Nov
- Hall 5 (AI & Robotics), Booth H5A033A
- BIEC Bengaluru
And on the last day, we’re hosting a Voice AI workshop. Details dropping soon!
See you there:)
Somil retweeted
🚨 Hiring Alert
Tag a designer. Win a trip to Singapore for you + a plus one ✈️
We're hiring a Lead Product Designer at @VaaniHQ
→ we are a foundation speech model company that also has a developer platform
→ the voice AI agents built on our platform are live on millions of CX interactions
→ you’ll design what users hear, how they build agents AND what they experience
→ chance to build your own team from scratch
→ 20-person team, HSR Bengaluru, full-time on-site
→ backed by Undisclosed, Venture Catalysts, @Google AI Accelerator & @nasscom
Know someone great? Fill the form below for us to consider as a valid referral👇
Please help spread the word. TIA.
The "I'm not a robot" checkbox doesn't check anything.
A bot can tick it just as easily as you. Websites have been quietly losing this fight for years, and now the bots have real browsers.
@PallaMohammad
I drew the whole story with stickmen 👇
mutedmiles.space/post/the-in…
Somil retweeted
Confession from all of us at Vaani: every Diwali, our phones turn into a to-do list.
We wondered what would happen if an agent did the calling. Very unscientific answer😅 47 calls, 50+ Indian languages and everyone gets wished.
(It’s a joke. Our agents usually call customers, not chachajis but we do love giving people their time back for the things matter, like family wishes and sweets.)
Wishing you and your loved ones a very warm and bright Diwali🪔✨
Team Vaani AI
a really well written blog
一个适合在个人电脑上学习微调+RL的小项目,虽然很小,但是是一个很好的学习案例
微调+RL Qwen3-1.7B
kevin-wu.ca/blog/qwn/
github.com/kevinyvv/qwen-3-1…
1
33










