Blog

An 11-language Indic voice-cloning model passed every infrastructure check we had — clean weight load, correct shapes, zero missing keys — and still came out as noise. Seven real bugs fixed correctly; the eighth was a checkpoint the vendor never actually published.

Every other model in this stack closed a language or capability gap. StyleTTS2 closed nothing — it was added purely for narration quality, and became the hardest deployment yet: a NumPy conflict traced to RunPod's own harness, a broken release, a PyTorch security default, and four packages with no wheel anywhere.

IndicWhisper gave OpenAudio Suite self-hosted transcription for exactly one language. SraVaani, shared as a link rather than a request, covers sixty-five — trained with a vision-alignment trick, deployed behind a license gate and an undeclared dependency, and tested on a language nobody in the stack could actually speak yet.

A pronunciation-control feature whose obvious \b-regex implementation passed a quick Hindi test and then silently matched nothing in Tamil, Telugu, Bengali, and Marathi — no error, just quiet failure.

Kokoro's 54 voices span nine languages — and exactly one of them is Indian. Closing that gap meant a natural-language-prompted model, a gated repo behind an Apache 2.0 badge, and two dependencies that have never published a wheel.

Every other engine had gone self-hosted. Transcription was still quietly calling OpenAI. Closing that gap meant a model that isn't on HuggingFace, a stale generation config, and a RunPod timeout that only bites on slow cold starts.

Two prior models proved the RunPod Flash playbook worked. This one broke it on the very first deploy — not because of RunPod, but because the model's own PyPI package couldn't be installed as published.

A real zero-shot voice cloning model, a torch version older than the GPU it landed on, and a RunPod storage backend that started failing at the worst possible moment. Five bugs were ours to fix. The sixth just needed a retry.

An 82M-parameter open-weight TTS model, a laptop with no Docker daemon, and five gotchas standing between "runs on CPU" and "a real user heard it in the browser."

A one-line static file, uploaded successfully by wrangler, present in the build output, byte-correct on disk — and unreachable in production. The Cloudflare config nobody warns you about until you ship a raw .html asset.

Four unrelated bugs — in deployment, design review, product logic, and a privacy disclosure — shared the same root cause: an assumption nobody had actually checked.

Three ways to give an AI coding assistant memory across sessions, tested on one small case: an internal accountancy project. The least automatic option won.