Rudransh🫟

Product Manager || AI & VOICE AI || I’ll post something daily.

[Rudransh🫟 @primepluto_]

Your accent is a label. Your idiolect is a fingerprint.
Same “Indian English” tag but your voiceprint and your friends are 512 dimensions apart.
The next layer of voice AI isn’t accent detection. It’s speaker-adaptive STT that learns you phoneme by phoneme, call by call.

[Shashi (シャシ) @shashpicious_]

If your design process still ends at figma, 2026 is going to hurt ⚠️

made designengineer(.)pro for designers ready to cross over before it’s too late

designing in cursor, shipping in code, AI in the loop, ditching the handoff. everything i wish someone handed me when i made the jump into it

Elevate your career from Product Designer to AI Design Engineer 📈🔥

Link Below 👇

[Rudransh🫟 @primepluto_]

Stared at the sky for 2 hours. Saw 3 Lyrid meteor and burn across Bangalore at exactly 4:05AM

[Rudransh🫟 @primepluto_]

Why did your voice bot give the wrong answer?
- Was it noisy audio? Wrong transcription? Bad intent classification? Hallucinated response?

- These are 4 different failures that look identical from the outside.

Built an MVP pipeline tracer that tells you exactly which domino fell first.

[Rudransh🫟 @primepluto_]

Yea Opus4.7 is cool, but this reassures me that I can always go back to GPT for most of my non specialised work without worrying about token.

[Claude @claudeai]

Introducing Claude Opus 4.7, our most capable Opus model yet.

It handles long-running tasks with more rigor, follows instructions more precisely, and verifies its own outputs before reporting back.

You can hand off your hardest work with less supervision.

[Rudransh🫟 @primepluto_]

I contributed to open source as a Product Manager !

Here’s what i found.📮

I ran Whisper-1 on 20 real Indian audio clips:

Pure English → 4% WER (error)
Pure Hindi → 16% WER (error)
Indian Names → 66% WER (error)
Hinglish → 100% WER (error) ❌

The 100% isn't the problem; the question is why.

The Hinglish finding isn't about mishearing.
Whisper switches the output script entirely romanized input comes back in Devanagari.
No flag exists to request Latin-script output for code-mixed audio. ( @SarvamAI wins here)
Blocker for any Indian support/WhatsApp transcription pipeline.

—
Dataset + Ground Truth: https://t.co/JtIwOmQrFs
Full Methodology: https://t.co/mlNs6t59iq

[Rudransh🫟 @primepluto_]

The hardest part of voice ai isn't the model.
it's silence.
what happens when the user pauses? thinks? changes their mind?
filling that gap well is the whole product.

[Rudransh🫟 @primepluto_]

Got voice AI latency under 670ms for $0.07/min without writing a single line of code.

The hack: using @Vapi_AI as a no-code inference playground.

Swapped ASR/LLM/TTS models in their UI, measure real latency + cost,
repeat.

current base pipeline:
Groq (Llama 3.1 8B) + hume_ai TTS.

groq’s speed makes up for llama’s weaker reasoning vs gpt-40/claude
and 670ms is right under the 1s threshold where conversations start feeling sluggish.

next: pushing for sub-500ms with Deepgram Nova-2.

[Rudransh🫟 @primepluto_]

Chat AI spoiled us.
Most people think Voice AI is just ChatGPT with a microphone. Wrong.

Every call runs through a pipeline:
VAD → STT → LLM → TTS → Audio
A chain of dependencies where milliseconds drop trust.

In Voice AI, speed is not a metric ; latency here is the product

Day 01/30 getting into Voice AI.

[Rudransh🫟 @primepluto_]

Claude this, Claude that. I’m doing things the way I always have. Not perfect but always out there.

[Rudransh🫟 @primepluto_]

@shashpicious_ Time to Upgrade to multiple monitors

[Rudransh🫟 @primepluto_]

Deciding between Claude code and the beaches 🧎🏻

[Rudransh🫟 @primepluto_]

@shashpicious_ 🤌🏼 always sharing the best deets

[Shashi (シャシ) @shashpicious_]

If your UI inspo starts and ends at Mobbin, you’re living in a small bubble 🫠

Here are some country-specific Mobbin (not really Mobbin, but you get it):

🇯🇵 Japan → https://t.co/RkZKUPnJsH

🇨🇳 China → https://t.co/dwQXnHyVno

🇰🇷 Korea → https://t.co/qSZXqPuSPr

Found these while designing a China-first app … completely different design brains at work.