IDEAS, DECISIONS & LESSONS

Journal

Notes on AI, architecture and building useful products.

21 articles

04 Mar 2026 · 13 min read

cost-optimization

Production Voice AI for Research at Scale: The Real Cost (Part 5 of 8)

Part 5 of 6 — Real-time per-minute cost tracking, provider comparison (OpenAI Realtime ~$0.053/min vs Gemini Live ~$0.029/min), budget enforcement with soft/hard limits, and the self-hosting math that saves 90% on transport.

02 Mar 2026 · 15 min read

data-pipeline

Production Voice AI for Research at Scale: From Recording to Insight (Part 4 of 8)

Part 4 of 6 — The 3-stage automatic pipeline that turns raw interview recordings into enriched, queryable research data in 3-7 minutes. Transcription, enrichment, analysis — with the transcript batching trick that cut DB load by 80%.

28 Feb 2026 · 12 min read

agents

Production Voice AI for Research at Scale: Multi-Phase State Machines (Part 3 of 8)

Part 3 of 6 — Research interviews follow structured protocols with distinct phases. How to build an LLM-driven state machine with next_phase() function calling and dynamic instruction swapping via set_chat_ctx().

26 Feb 2026 · 13 min read

debugging

Production Voice AI for Research at Scale: Zombie Agents, Pre-Warming, and the 5 Bugs That Cost Us Weeks (Part 2 of 8)

Part 2 of 6 — The production pain points nobody warns you about: zombie agents, metadata latency, pre-warming for 1-2s time-to-first-voice, VAD tuning for research respondents, and provider quirks.

24 Feb 2026 · 11 min read

architecture

Production Voice AI for Research at Scale: The Architecture Nobody Warns You About (Part 1 of 8)

Part 1 of 6 — Why research interviews need server-side voice agents, the three-tier architecture, room metadata as configuration transport, and the 100-500ms propagation latency nobody tells you about.

23 Feb 2026 · 28 min read

agents

The Voice AI Interview Playbook: Multi-Role Agents — Interviewer, Coach, and Evaluator Personas (Part 5 of 12)

Part 5 of 12 — Three personas, one infrastructure. How to build an AI interviewer that asks questions, a coach that gives feedback, and an evaluator that scores fairly — all with system prompts, function calling, and state machines.

22 Feb 2026 · 22 min read

The Voice AI Interview Playbook: Building the Voice Pipeline — STT, LLM, and TTS That Actually Work (Part 4 of 12)

Part 4 of 12 — Deepgram Nova-3 or Whisper for STT? Gemini Flash or GPT-4o for conversation? ElevenLabs or Cartesia for TTS? Real benchmarks, real costs, and the combinations that actually work in production.

21 Feb 2026 · 18 min read

livekit

The Voice AI Interview Playbook: LiveKit vs. Pipecat vs. Direct — Picking Your Framework (Part 3 of 12)

Part 3 of 12 — LiveKit gives you WebRTC infrastructure. Pipecat gives you pipeline flexibility. Direct integration gives you simplicity. Here's how to choose for your voice interview platform.

20 Feb 2026 · 29 min read

gemini-live

Building a Production Voice AI Interview Agent: S2S-Only with Dynamic Prompts

A practical build guide for a speech-to-speech interview agent using LiveKit MultimodalAgent, OpenAI Realtime, and Gemini Live. Dynamic system prompts, 3 personas, function calling, and provider switching — no cascaded pipeline needed.