Skip to content
Valtair
Insights

Field notes from building AI products

Technical articles, product build logs, and analysis for people building with AI. Depth over daily news.

Every piece here comes from designing, building, and running AI systems: what worked, what did not, and the decisions behind it. It is written for engineers, founders, and product teams past the demo stage who need things to hold up in production. Start with a topic below, or subscribe to the newsletter to get new pieces as they publish.

Voice and conversational AI

How to build voice agents callers do not hang up on: the architecture, latency and turn-taking, handing off to people, and evaluating an agent before it takes real calls.

Explore Voice AI
Cascaded or speech-to-speechSeparate speech-to-text, language model and text-to-speech stages with a transcript are compared with one speech-to-speech model. Both carry call audio over telephony.Cascaded pipelineSpeech to textTranscriptModelText to speechSpeech-to-speechTelephony
Article7 min read

Voice AI agent architecture: from phone line to response

The layers of a production voice agent, what each one does, and how to choose between a cascaded pipeline and speech-to-speech.

Voice and conversational AI
Voice / human handoffAn editorial illustration of a conversation branching to a human, with its context carried along.ConversationTool executionYour teamContext
Article4 min read

When a voice agent should hand off to a person

The signals that a call needs a person, how to transfer context with the call, and how to measure the balance.

Voice and conversational AI
Voice / evaluationAn editorial illustration of call scenarios being tested, reviewed, and fed back into evaluation.Call scenariosSimulateHuman review
Article4 min read

How to evaluate a voice agent before it takes real calls

What to measure, how to build and simulate call scenarios, and how to roll a voice agent out safely.

Voice and conversational AI
Voice / timingAn editorial illustration of a streaming voice pipeline: speech recognition, the model, and speech synthesis overlap in time, so the first audio goes back to the caller before the full reply is ready.Speech to textModelText to speechCaller stopsFirst audio back
Article7 min read

Designing voice agents for latency people can live with

Where the seconds go in a voice pipeline, and the architecture choices that get them back.

Voice and conversational AI

RAG and knowledge systems

Building systems that answer from your own data with citations, and measuring whether those answers are actually supported by their sources.

Explore RAG and Knowledge Systems
Grounding / evidenceAn editorial illustration of the claims in an answer being checked against retrieved sources, with one claim that no source supports.Answer claimsNo sourceClaims checked against sources
Article3 min read

Evaluating grounded answers in RAG systems

A practical approach to measuring whether a retrieval system is actually answering from its sources.

RAG and knowledge systems

Product build logs

Engineering notes from building Valtair's own products: the decisions, the trade-offs, and what we would do differently.

Explore Agents and Automation
Routing / reuseAn editorial illustration of task-based model routing and a cache that reuses work, without presenting numerical cost results.Research taskRoutine workDeeper reasoningReuse cached work
Build log3 min read

How we kept LeadVector's inference costs predictable

The routing, caching, and evaluation decisions that stopped model spend from scaling with usage.

Product build logs
Valtair Signal

Practical intelligence for people building with AI.

A concise briefing on important AI developments, product engineering lessons, emerging use cases, and what technical and product teams should do next.

Twice per month. No noise. Read our privacy notice.

Book a discovery call