
Per aspera, ad astra Building @AvarthaAI | ex SiriAI @Apple | CS @GeorgiaTech
Cupertino, CA
Joined October 2018
- Tweets80
- Following416
- Followers188
- Likes1.1K
A few months ago I left @Apple Siri AI to build with the most talented systems for AI team I ever worked with. Happy to bring this to fruition this soon.
We are just getting started 🚀
Introducing Avartha.
We're building the infrastructure to bring frontier-grade intelligence to real-time applications. Our first release, Ava Zero, is a voice agent entirely running open models using our inference infrastructure and ranks #1 on Sierra's tau-voice benchmark with highest task completion rate and generation quality and just costs $2.5 per hour.
Open models are now at the frontier of voice AI.
Asmit retweeted
i was demoing @usedayly and the person cancelled subscription for my competitor in front of me
Ava doesn’t just handle support conversations, but it can handle very complex workflow orchestrations.
Internally, Ava controls codex, @claudeai , @opencode sessions for us and just gets our work done.
Ava improves herself technically.
Voice is poised to be the dominant interaction surface for all of computing in the future. So far voice ai have been held back by chunky models, but when you plug in models like @Zai_org's GLM 5.3 in the loop, possibilities become endless.
Here is a snippet of what our team has been cooking with Ava, our voice agent. Happy Halloween! 🎃
Congrats to the @AvarthaAI team for topping the Tau Voice 2.0 leaderboard! This means they win against the big labs across 5 languages, on not just task completion but also speech fidelity and multilingual naturalness: tau is more robust now! And all using open weight models.
Introducing Avartha.
We're building the infrastructure to bring frontier-grade intelligence to real-time applications. Our first release, Ava Zero, is a voice agent entirely running open models using our inference infrastructure and ranks #1 on Sierra's tau-voice benchmark with highest task completion rate and generation quality and just costs $2.5 per hour.
Open models are now at the frontier of voice AI.
Asmit retweeted
Five months ago, we started Avartha with the belief that better systems can unlock the next generation of real-time applications. Today, we’re announcing Ava Zero, our voice agent built entirely on open models. Ava ranks #1 on @SierraPlatform voice agent benchmark, ahead of OpenAI.
Incredibly proud of our team. We have some of the best systems builders in the world, and we’re just getting started.
Introducing Avartha.
We're building the infrastructure to bring frontier-grade intelligence to real-time applications. Our first release, Ava Zero, is a voice agent entirely running open models using our inference infrastructure and ranks #1 on Sierra's tau-voice benchmark with highest task completion rate and generation quality and just costs $2.5 per hour.
Open models are now at the frontier of voice AI.
Asmit retweeted
. @Uber_India @UberIN_Support a perfect example of pot calling kettle black today. My friend and I were using uber while getting out at our destination, we politely requested the driver to drop us inside the society because it was raining and we had heavy luggage. He refused very rudely and started behaving aggressively and asked to pay him and refused to open the trunk of the car to get the luggage out and said nonsense like you are the kind who don’t pay etc. Then he started behaving very aggressively towards my female friend and started to get inside the car and accelerated his car towards her. The guards of the society saw the altercation and came to the rescue. They requested him to park the car aside, so they can talk to him, but he quickly reversed the car and fled.
We reached home and reported the incident immediately, and the customer service representative was very empathetic and assured us that action would be taken immediately. But then, in the evening, we have received two contradictory emails from Uber, which are attached. Kindly resolve this as it is a safety issue of the highest order, and my friend is afraid of using any cabs or even getting out of her house because he knows her location.
my number one favorite feature of new siri is being able to ask for clarity & action on my screen context.
this is akin to how valuable it is to take screenshots as a feature.
only the os can provide this type of a simple user interface dynamic to leverage intelligence on demand with as little friction as possible. it’s a joy to use. i hope at some point i am able to tune it a lot more & have a choice on using the models i love even for personal context.
Asmit retweeted
I've been using the new Siri on Mac by way of the Terminal Command workaround (still not off the waitlist). My take so far: Apple nailed it. It's awesome for a beta 1. (Hopefully this gets me off the waitlist but it is true).
Presented our work from Apple at EuroMLSys 2026 in Edinburgh:
“Asynchronous Verified Semantic Caching for Tiered LLM Architectures”
The core idea:
semantic caches are usually conservative because false positives are catastrophic. Instead of pushing more verification into the serving path, we moved verification off the critical path entirely. Near miss cache interactions asynchronously trigger an LLM judge verifier. Verified pairs are promoted into the dynamic tier over time, increasing effective cache coverage while preserving the latency behavior of the original system.
Interesting systems tension:
you want higher cache hit quality without paying synchronous verification costs on user requests.
Paper:
machinelearning.apple.com/re…
#MLSys #LLM #Inference #Caching #EuroSys
4
323
Asmit retweeted
Incredibly excited to release GPT-5.1 Pro to the world! It’s great at tackling the hardest, messiest problems with clearer, more comprehensive responses.
I’m eager to see you throw your toughest problems at it. Please try it and share what you think :)
Asmit retweeted
The bitter lesson of AI infra: The hardest part about building faster LLM inference systems is not designing the systems, but rather it is evaluating if the system is actually faster! 🤔
This graph from a recent top systems venue paper about long-context serving shows average normalized input token latency for a trace with both short and 100K+ token requests. System X looks like a clear win: lower normalized latency and higher request rates. But normalized metrics can obscure the actual user experience: at those rates, long inputs see >2hr delays to the first token!
Let’s do the math!🧮
Asmit retweeted
Super long-context models with context window spanning millions of tokens are becoming commonplace (@GoogleDeepMind Gemini, @xai Grok 3, @Alibaba_Qwen Qwen2.5). But efficiently serving these models is tough, especially alongside short requests. Head-of-Line (HOL) blocking becomes a major issue, hurting latency for everyone.
We present Medha, a system designed to handle this mix efficiently. Achieving 30x lower latency, and 5x higher throughput compared to the state-of-the-art. Full paper: arxiv.org/pdf/2409.17264. 🧵
AGI will be the ultimate monument to human pride, and simultaneously its ultimate undoing
99.99% of people cannot comprehend how insane FrontierMath is.
The problems are crafted by Math profs and not in any training data.
Math legend Terry Tao said "These are extremely challenging. I think they will resist AIs for several years at least."
OpenAI o3 did 25% on THIS.
4
894
This Apple Intelligence feature hits the spot for inquisitive but impatient types like me. Information snacking.
5
881






















