Hamming AI

Hamming AI is now SOC 2 compliant. 🎉 This milestone reflects something core to who we are: helping companies test, monitor, and deploy safe, production-ready voice agents. That’s only possible when the underlying systems are secure, reliable, and enterprise-grade. Achieving SOC 2 demanded months of work across engineering and operations, from continuous CVE monitoring and patching to building incident-response playbooks and tightening every layer of our infrastructure. For our customers, this means even stronger guarantees around data protection, availability, and operational excellence. Security isn’t a feature. It’s the foundation. And we’re just getting started.

We're excited to share that we've successfully completed our SOC 2 audit, achieving full SOC 2 compliance. From day one, we’ve built Hamming on the belief that reliability isn’t a feature, it’s a foundation. SOC 2 validates that the systems behind our platform meet rigorous industry standards for security.

Over the past few months, we’ve been speaking with a lot more healthcare companies, triage teams, care-coordination groups, digital health platforms, all trying to roll out voice agents safely. The stakes are higher when deploying voice agents in clinical settings, not just cause of HIPPA and potential PHI leaks, because when a voice agent fails in a clinical workflow, the failure can directly put a patient at risk. For Hamming AI's blog, I wrote about the five failure modes that make voice agents unsafe in clinical settings. Link in comments! TL;DR: The five failure modes are: 1. Perception failures 2. Guardrail failures 3. Multi-agent reasoning failures 4. Workflow logic drift 5. Latency & infrastructure drift If you’re working on clinical voice agents and have any questions, let me know!

I’ve been thinking a lot about accuracy lately. In conversational AI, accuracy is the metric that tends to get the most attention. But the more time I spend with voice agents, the more I’ve realized: Accuracy captures only a thin slice of how voice agents actually perform. A voice agent can maintain a high accuracy score while drifting across turns, skipping guardrails, or responding to an ASR hallucination that never existed in the audio. And none of that shows up in a simple pass/fail metric. This came up in a previous episode of The Voice Loop (our podcast) with Fabian Seipel from ai-coustics. The audio layer alone introduces distortions most teams never evaluate: clipped speech, mic inconsistencies, compression artifacts, overlapping voices, etc. The reality is: Accuracy measures the endpoint and not the behavior of the pipeline that produced it. But conversational AI is a pipeline problem: ASR → retrieval → reasoning → response. If any of those layers misinterpret, drift, or break, the “accurate” answer can still be wrong in practice. I wrote more about my thoughts on the Hamming AI blog. Link in comments!

Testing how voice agents perform in both real-world audio environments with natural, imperfect human speech is the combo that gets us closer to a true, Turing-level benchmark, because it reflects how people actually communicate. 🎙️ The Voice Loop Episode 2 is out now. Link in comments

When you're building in voice AI, you'll realize audio quality is one of the most underrated layers in the entire stack. When you analyze audio at scale and dig into thousands of real user recordings, you start seeing distortions you never planned for, not just background noise or bad mics, but strange, hyper-specific issues that only show up in real environments. Every new dataset reveals something new. Teams like ai-coustics do something counterintuitive: They intentionally destroy clean audio so models learn to survive messy, real-world conditions. And that’s exactly why, at Hamming, you can simulate the environments your voice agents will actually face, from cars and kitchens to drive-thrus and call centers, etc. If production variables are unpredictable, your testing variables have to be too.

Most people don’t realize how much raw complexity sits inside a single audio signal. In our latest episode of the Voice Loop, Sumanyu Sharma and Fabian Seipel unpack a highly technical component of audio engineering: source separation. The ability to take one audio track and break it into its individual components. Back in 2014, Fabian’s internship at Sony sparked his curiosity. Here's what he had to say 👇️