Kuzushi Labs · Engineering notes

Engineering conversational AI

From customer requirements to a dependable telephone conversation. Twelve articles on the decisions behind the client platform built with Bolna, Zoice built with Pipecat, and the separate 2026 speech-model work.

Arunank Sharan · July 2026 · 12 articles

1. PART 01 · 12 min read

   Conversational AI Design: the call is the product

   From customer requirements to an audio, reasoning and business-action system.

2. PART 02 · 11 min read

   Why an 800 ms voice response is a systems problem

   An explicit latency budget, turn detection, cancellation and playback correctness.

3. PART 03 · 10 min read

   A Bolna-based telephony agent, from connection to hangup

   Call lifecycle, asynchronous queues, provider adapters and production boundaries.

4. PART 04 · 10 min read

   Designing a configurable voice platform with Pipecat

   Frames, lifecycle, SEDA principles and interoperable cascade and native-audio paths.

5. PART 05 · 10 min read

   Choosing the speech and reasoning models for a voice agent

   Provider trade-offs, native audio, language coverage and a workload-specific Pareto frontier.

6. PART 06 · 10 min read

   Self-hosted Whisper: streaming, scheduling and the concurrency trap

   Rolling recognition, stable prefixes, GPU queues and a capacity calculation for real calls.

7. PART 07 · 13 min read

   Adapting a 2026 speech recognizer for English, Hindi, Singlish and Hinglish

   A Qwen3-ASR reference experiment, plus an honest analysis of a completed Whisper Tiny run.

8. PART 08 · 16 min read

   Speech synthesis that can survive a real customer call

   Pronunciation, phrase streaming, a self-hosted Indic TTS candidate and adaptation protocol.

9. PART 09 · 12 min read

   Context, memory and tools: where a voice agent becomes a business system

   Conversation state, retrieval, streaming function calls, idempotency and interruption-safe actions.

10. PART 10 · 20 min read

    Running voice agents on AWS and GCP without losing the conversation

    Cloud topology, Bedrock, GKE/EKS, GPU serving, admission control and graceful draining.

11. PART 11 · 19 min read

    Evaluating a voice agent: listen to the call, inspect the action

    A practical evaluation system across audio, interaction, policy and verified business outcomes.

12. PART 12 · 11 min read

    What a voice agent changes for the business

    Unit economics and business outcomes from the client platform and Zoice.

Related case studies: the client calling platform and Zoice.