01 · Data sovereignty is still the #1 reason enterprises self-host A global enterprise in Taiwan self-hosts everything on Dograh because they won't let call data leave the country. An Australian BPO wanted full air-gapped, on-prem GPUs. They wouldn't even share a call script with ChatGPT. Compliance isn't a checkbox for these teams, it's the whole decision.

02 · Self-hosting isn't cheaper, it's differently priced Running a voice agent fully on-prem at ~5 concurrent calls costs roughly $500/month: about $300 for a GPU running a small model, $150–300 for the orchestration server. The real cost isn't the infra, it's the DevOps effort to keep it running, something almost everyone underestimates.

03 · There's a real volume threshold for self-hosting Below it, don't bother. Dograh's rule of thumb: self-hosting starts to make sense around 100K–200K calls a month. Below that, or if speed to market matters more than control, a cloud platform wins every time.

04 · Sometimes it's not a logical decision, it's an emotional one "It's like buying a house instead of renting. It doesn't always make financial sense, but people do it anyway." Pritesh described a Florida restaurant chain whose founder self-hosted against their advice, purely because he loves building. Passion for owning your stack is a real, recurring driver.

05 · The reliability math nobody talks about Multiply it out: ~80% transcription accuracy × ~90% LLM accuracy × ~90% speech generation accuracy ≈ 63% odds of a clean turn. Over a 10-turn conversation, that's single-digit odds of a flawless call. It's the clearest explanation yet for why voice AI feels so much harder to get right than it looks.

06 · The middle ground: data-boundary hosting, not full air-gap Most enterprises don't need to fully air-gap. Hosting models inside a cloud provider's VPC or data perimeter keeps data from ever leaving the boundary, satisfying compliance without the cost and complexity of running GPUs in a basement.

07 · Post-launch is where the real surprises show up Mai shared a real estate voice agent built to answer listing questions, but once live, users mostly asked whether a neighborhood was safe or good for schools. The agent had no context for that, so it hallucinated confident-sounding answers instead. The gap between what you build for and what people actually ask is where most failures start.

08 · Being open source makes them faster, not just cheaper Dograh's users push PRs and flag gaps directly in their community: feedback closed platforms simply don't get. Pritesh's blunt (and funny) answer for why they went open source in the first place: "we wanted to watch the world burn." The real reason is more standard OSS conviction, but the line got a laugh for good reason.

09 · Even a fully open-source platform didn't build its own observability Dograh ships as a complete stack (database, file storage, cache, its own MCP), but for monitoring, they integrate Tuner rather than building it themselves. Abhishek: you don't fiddle with SDKs, "you literally go and drop a box, create an account on Tuner, configure the credential, and you're set."

10 · Speech-to-speech vs. cascade isn't settled, it's use-case specific Speech-to-speech wins on multilingual conversations and latency (Gemini's live models handle 70+ languages interchangeably). Cascade wins when you need tight control, like outbound calls with strict logic. Pritesh's take: build for speech-to-speech if you're building for where the industry is headed.