The question I probably get asked the most: how do you know your voice agent is actually working in production? I think of it like blood work. One marker tells you very little. Grouped, they show a pattern, and the pattern points to what needs fixing. Voice just has more layers.

Is there a GTM tool where you can define your own signals and it just gives you leads based on them? And I don't mean the standard “raised a round, hiring SDRs” stuff. Signals specific to my market. What are you all using?

Never imagine I would come to speak at the college my dad went to. Manipal Institute of Technology :) #TedX Excited for tomorrow ♥️ #MIT

We mapped 16 voice AI companies building in automotive retail. Then we pulled the 5 failure modes we see most in live dealership calls at @SuperBryn_AI. Most of them sound fine on the call. The damage shows up later. Full company list in the thread blow. Bookmark this and let me know if we missed someone.

Building something of our own, together. Gandhi Jayanti spent well 🙏 #theusualSuperBrynshenanigans @devanshtrip @parthjain_1 @DrNeethuJoy

A lot of you have found me here over the last couple of weeks, so probably a good time to introduce myself properly :) I'm Nikkitha, co-founder and CEO of @SuperBryn_AI. We're building the reliability layer for voice agents, and these days I split my time between Bangalore and SF. A few things on how I got here. > Was born in Kerala and grew up in the Middle East. > After graduating, I decided I wanted to become a civil servant and spent 3 years preparing for the UPSC. (failed pretty miserably, btw 😅) > During my final year of preparation, I joined a startup, and somewhere along the way, I fell completely in love with the world of startups. > Eventually started Shoppre, an ecomm cross border tech comapany and grew it to ~$2M in annual revenue. > Shoppre was profitable and scaling, but I knew I wanted to build in AI. So I stepped away from operations (the team runs it beautifully on their own) and chose to start from 0 again. > For almost a year, @DrNeethuJoy and I went through the whole 0→1 mess. > After a lot of iterating, that became SuperBryn. > Raised $1.2M from @Kalaari and angels, and today we work with companies around the world on making their voice agents reliable. > Outside work, I like to dance, make art when I get the time, and give an unreasonable amount of attention to my dog, Django :) If you're new here, hello 👋 I mostly write (or at least try to) about what I'm learning while building SuperBryn: Voice AI, enterprise AI, startups, customers, and occasionally the strange career decisions that got me here.

On Saturday at TGPF, we had almost the whole voice AI stack on one stage 😅 Telephony. Inference. TTS. Evals. And people who’ve actually shipped voice AI across markets. What surprised me was how quickly a conversation about architecture became a conversation about GTM. > 0→1 in voice is as much about the network you build as the product you ship. > Founder-led content, events and partnerships aren’t side projects either. They’re often where land-and-expand starts. > And then there’s the product economics: every second of a call costs money, infra choices affect margins, and > what enterprises ask for in a demo isn’t always what they eventually pay for. My part was on production readiness. The analogy I keep coming back to is blood work. You don’t really judge health by how someone looks, do you? You track markers and patterns. Voice AI needs the same thing across telephony, STT, LLM, TTS and orchestration, tied back to outcomes like resolution, conversion and escalation. And because these systems are probabilistic, evals can’t be a final checklist. If you’re building the agent, you’re building the eval harness alongside it. Otherwise, today’s fix is most likely to become tomorrow’s regression I really enjoyed being a part of this panel with Suman, Devansh, Sahil, Jagriti 😄 And Swati, Amrit and the TGPF team truly knocked it out of the park with the event 🙌

We tested Jev ourselves: 0.55s median latency from India, too slow to check every turn of a live call. So where does it fit in voice AI? Risky actions mid-call, screening every call after, catching fixes that break what worked. And where it doesn't.