Harshit Mishra

So i'm building a Voice AI platform and, Our knowledge base lookup latency is ~20ms. Most voice AI platforms don't publish this number. Some don't even measure it. Here's exactly how we got there with actual logs. Thread.

@ardent__dev verblyai.com Where Conversations Become Operations. We are verblyai a voice ai platform that drives growth.

Founders 👇 Drop your product below. Let's drive some traffic your way 🚀 The best ones will get featured on EverFeatured a curated, quality-first product directory.

@mscode07 verblyai.com Where Conversations Become Operations. we are verbly.

Drop your product 👇 Let's do some Marketing!!

@ionleu verblyai.com we are a voice ai platform similar to vapi

drop ur startup link

@csaba_kissi Hey this is mine , I'm building a voice ai platform similar to vapi. verblyai.com

Share your website/project, guys👇

@iarchitX @suni_code Hey, it's look interesting

@suni_code I’ve created a AI Jailbreak Playground. Do try this and post your scores: ai-jailbreak-challenge.vercel.app

Drop your project URL 👇🏻 Let’s drive some traffic!!!...

@maurikmillaku @suni_code Looks clean and minimal

@suni_code This is just my web: maurikmillaku.com

@suni_code Hey this is mine, I'm building voice ai similar to Vapi. verblyai.com

pre-warming the index at call connect dropped average KB latency from ~180ms to 23ms. Not a model change. Not more compute. Just: do the work earlier. Most latency in voice AI is avoidable with architecture, not hardware.

What we're running under the hood: → Qdrant for vector search (self-hosted, same region as inference) → Embedding pre-computed at agent config time, not at call time → Retrieval triggered on partial STT transcript (first 70% confidence threshold) → Top-K=3

The fix isn't a faster vector DB. It's architectural: → Pre-warm the KB index at call start, not at query time → Run retrieval async kick it off the moment STT begins transcribing, before the transcript is even complete → Cache the top-K chunks from the first 3 turns

First: why KB latency matters more than people think. In a live voice call, every millisecond of silence feels like 10x that to the caller. If your KB lookup takes 400ms synchronously during inference, you've just added 400ms of dead air. The user thinks the agent is broken.

Here's what a naive KB implementation looks like in the logs: STT transcript received LLM inference started KB lookup triggered ← inline, blocking KB result returned ← 383ms gap LLM first token out TTS synthesis started 383ms of silence. On a voice call. Unacceptable.

@aditiitwt Hiii

@LewisWnLP CRO

I've compiled 110+ Landing Page Before/Afters in a single file. This Figma file is yours for FREE. Like + Comment "CRO" and I'll send it to you (must be following) RT is appreciated

"Congratulations to the champions of FLASH HACK'25! Your innovation, dedication, and teamwork have earned you the winner's title. Keep coding, keep creating! #FlashHack25 #Winners #GoogleDeveloperGroups #RCETBhilai"

Announcing Our Finalists! 🎉 After rounds of hard work and innovation, these teams have made it to the finals! Stay tuned for the final showdown! #FlashHack #GDGRCET #RungtaCollege #GoogleDeveloperGroups

These are the core problem statement categories for the Google Solution Challenge. #GDG #solutionchallenge #gdgoncampus

What are you waiting for ⁉️ Assemble your team, brainstorm, and get those creative juices flowing. Register now :-shorturl.at/wERiw ‼️ #AIforSocialGood #SolveWithAI #SolutionChallenge2025 #GDGSolutionsChallenge