HowwewalkedintoYC

This isn’t a victory lap. It’s the honest version of how a voice AI idea became something companies pay to run their phone calls on, and how that turned into a YC acceptance. I was the first founding engineer at Samora, and I built the spine of it: the voice pipeline, the infrastructure, and the retrieval layer that keeps it accurate. So this is told from inside the repo.

Before Samora was a company with a YC batch, the team had already learned the hard version of this problem: scaling a voice agent to thousands of users across India as part of a national helpline. That’s where the romance of “AI voice agents” met reality.

Linguistic diversity. People switching languages mid-sentence. Dialects the models had never really heard. Choppy lines, background noise, networks dropping. Local infrastructure constraints. It exposed, very quickly, the gap between a demo that works on stage and a system that works on a Tuesday afternoon for someone on a bad connection. Reliability, observability, and controlled escalation stopped being nice-to-haves and became the entire job.

Phone calls still power core workflows: support, collections, verification, recruiting, outreach, feedback. Voice AI platforms made it easier to build an agent. But once it’s live, someone still has to manage runtime behavior, watch the edge cases, enforce policy, investigate failures, and handle escalation when a conversation turns sensitive.

Here’s the insight that became the company: most teams do not want to build an internal AI voice operations function. They want the outcome, not the on-call rotation. So Samora operates the agents in production for them, and owns deployment, monitoring, compliance guardrails, observability, and human escalation.

Multilingual, domain-aware voice agents that handle real inbound and outbound calls, and the operational layer that keeps them trustworthy:

Code-switching and dialect-heavy conversations. Strict policy enforcement and compliance controls. Integrations into CRMs, ATSs, and ticketing. Turn-level observability and structured explainability. And managed human escalation, with trained Samora operators stepping in when a conversation gets uncertain or sensitive.

Every call is fully logged and auditable. A customer can see what the agent heard, which rule fired, what action it took, and exactly why it escalated. That audit trail is the difference between a clever demo and something a regulated business will put on its main line. Hover a stage below to see how a single turn flows through the pipeline I built.

A lot of teams can get a voice agent to say a sentence nicely. Very few can stand behind it in production across financial services, recruitment, healthcare, consumer goods, automotive, and government workflows, which is exactly where Samora’s paying customers are.

The proof was in the numbers, not the pitch: 80% first-contact resolution in support use cases, a 12% lift in collection rates for an auto-finance company, and 100,000+ candidates reached in recruiting while saving 1,700 recruiter hours. Most pilots go live in under a week.

That’s what makes the technology defensible. The hard part was never the model, it was everything around it: making the system reliable on bad lines, observable at the turn level, compliant under real policy, and safe to hand a sensitive call to a human at exactly the right moment. We’d already paid for that knowledge on the national helpline. YC was buying a team that had survived the production version of this problem.

So back to the conviction. When I told the founders we’d get into YC, it wasn’t bravado, it was math on what we’d built. We weren’t selling a hope. We were selling a system that already worked under the worst conditions we could find, with metrics from real customers and an operational story most “voice AI” companies can’t tell.

I’d built the pipeline from an empty repo into something carrying serious daily volume. I’d watched it hold up when the lines were bad and the dialects were hard. I knew what was under the hood. That’s why I could say “you’ll see, we’ll get in” and mean it.

And we did.