VoiceFlow: How a 2-Minute Voice AI Became a 3-Second One (and Won Nationally)

Jitesh Kumar — Fri, 04 Sep 2026 13:31:33 GMT

I built VoiceFlow because of one annoying, very specific problem: sales reps finishing a call and then manually re-typing everything that was just said into a CRM. Name, intent, follow-up date, all of it, typed in by hand, every single call. Real estate firms in particular lean on this kind of outbound calling heavily, and it usually means hiring a marketing and sales team just to keep the pipeline moving and the CRM updated. I wanted to see if a voice AI agent could do both at once, talk to the lead, and update the CRM live, mid-conversation, with no human re-typing anything afterward.

That's what VoiceFlow does. It's a full-stack real-time voice agent, React.js on the frontend, FastAPI on the backend, that integrates with Salesforce Agentforce to push CRM updates the moment they happen in the call, not after. For the actual conversation, Azure OpenAI handles the reasoning and Murf AI's API handles the voice.

The version that worked but shouldn't have shipped

The first working version of VoiceFlow was, technically, a voice AI agent. It just wasn't a usable one. It took two to three minutes to respond after someone finished speaking. On a phone call, two to three minutes of dead air is not a pause, it's a hang-up. The pieces were all there, speech understood, intent extracted, CRM updated, but strung together sequentially: wait for the LLM, then wait for the voice synthesis, then send it back. Every step blocked the next one.

Re-engineering the pipeline

The fix wasn't a smarter model, it was a smarter pipeline. I rebuilt the audio pipeline around Python multithreading and async WebSocket streams so the different stages of a response, reasoning, CRM writes, voice synthesis, could happen concurrently instead of waiting on each other in a queue. Backend response latency dropped 70%, down to under 2 seconds, and the full round-trip a caller actually experiences, from finishing their sentence to hearing VoiceFlow's reply, came down to somewhere between 3 and 10 seconds depending on how much the agent has to reason through.

That's the difference between a demo and something you could plausibly put in front of a real caller.

Murf AI Coding Challenge 5

I built VoiceFlow solo and entered it into Murf AI Coding Challenge 5, starting around September/October 2025. It won 1st Rank nationally, against hundreds of participants, one of five national hackathons I've won so far, and the one that ended up mattering the most, because it's what led to the paper.

From project to published research

Somewhere in the process of explaining why the latency fix worked, I realized I had the shape of an actual research paper, not just a project writeup. That became "AI-Powered Voice Agent System for Business Communication: A Latency Optimization Approach", first-authored, presented at the 2026 IEEE International Conference for Convergence in Computing Technology (I3CTCON), IEEE Maharashtra Section, in March 2026.

I still think about VoiceFlow and what else could be pulled apart and rebuilt in it. The latency fix solved the problem that made it unusable, not the last problem it has. That part hasn't stopped.

Code's on GitHub if you want to look under the hood: https://github.com/JITESH-KUMAR05/VoiceFlowAI

This is the first piece in my tech-journey series, next up: what three months as an SDE intern taught me, and the story behind a MSME-funded student project.