--- title: Tolan: Voice-First AI Companion — Paula Dozsa, Tolan description: An engineer from Tolan, a voice-first AI companion app, describes the architectural challenges of building real-time voice conversations with LLMs:... canonical: https://daily.dev/posts/tolan-voice-first-ai-companion-paula-dozsa-tolan-bvu47mw5k twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Tolan: Voice-First AI Companion — Paula Dozsa, Tolan | daily.dev og:description: An engineer from Tolan, a voice-first AI companion app, describes the architectural challenges of building real-time voice conversations with LLMs:... og:url: https://daily.dev/posts/tolan-voice-first-ai-companion-paula-dozsa-tolan-bvu47mw5k og:image: https://api.daily.dev/og/posts/bVU47mW5k.png og:image:alt: Tolan: Voice-First AI Companion — Paula Dozsa, Tolan og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Tolan: Voice-First AI Companion — Paula Dozsa, Tolan **[AI Engineer](https://daily.dev/sources/aidotengineer)** · 15 min read · 0 upvotes · 0 comments ## Summary An engineer from Tolan, a voice-first AI companion app, describes the architectural challenges of building real-time voice conversations with LLMs: sub-2-second latency budgets, smart turn-taking to handle interruptions, a tiered model routing system (frontier models for high-stakes moments, cheaper models for casual chat), a retrieval-based memory system with nightly compression, and per-turn context rebuilding rather than cache reuse. The talk also covers how the team uses Claude-based agent fleets (implementation, review, PR shepherd, triage agents wired into Linear, Sentry, and Datadog) to build the iOS app itself, reporting improved crash-free rates and reduced runtime errors, plus a hiring pitch for engineering roles including a new agent-management position. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How does Tolan decide which LLM to use for a given conversation turn? Tolan runs a tiered model fleet and routes turns based on emotional stakes rather than cost. A lightweight classifier called the tone router, running on a cheap model, reads the emotional state of each turn; high-stakes moments like a user's first message or crisis-level conversations always get the frontier model, while casual back-and-forth and background tasks run on smaller, faster models that cost roughly five times less per turn. _Teams weighing frontier versus small-model routing for cost control can track real-world architecture patterns like this on daily.dev._ ### What made the biggest improvement in voice AI response latency for Tolan? Switching to GPT-5.1 on the responses API cut time to speech by more than seven-tenths of a second, the single biggest latency improvement reported. This mattered because the full pipeline (end-of-utterance detection, transcription, first token, text-to-speech, playback) needs to complete in under about two seconds total, since a half-second increase in latency measurably hurt user retention metrics. _Developers chasing real-time LLM latency budgets can follow model-swap wins like this one on daily.dev._ ### How does an AI companion app handle long-term memory without blowing up context length? Rather than storing full conversation transcripts, Tolan treats memory as a retrieval system: facts, preferences, and emotional signals are extracted, embedded, and stored in a vector database with sub-50-millisecond lookups. Nightly jobs merge duplicates, cluster related memories, resolve contradictions, and drop noise, while memory is split into a volatile live-tail and a stable, cacheable block based on what actually gets recalled. _Anyone designing memory systems for conversational agents can track retrieval-based approaches like this on daily.dev._ ## Similar posts on daily.dev - [How I Made My Own Offline Voice Assistant](https://daily.dev/posts/how-i-made-my-own-offline-voice-assistant-lvephkfkw) · Medium · 0 upvotes · 0 comments - [How I built a sub-500ms latency voice agent from scratch](https://daily.dev/posts/how-i-built-a-sub-500ms-latency-voice-agent-from-scratch-neao8lbjr) · Hacker News · 0 upvotes · 0 comments --- Tags: [#ai-agents](https://daily.dev/tags/ai-agents), [#claude](https://daily.dev/tags/claude), [#voice-ai](https://daily.dev/tags/voice-ai) [View this post on daily.dev](https://daily.dev/posts/tolan-voice-first-ai-companion-paula-dozsa-tolan-bvu47mw5k) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Tolan: Voice-First AI Companion — Paula Dozsa, Tolan","url":"https://daily.dev/posts/tolan-voice-first-ai-companion-paula-dozsa-tolan-bvu47mw5k","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/tolan-voice-first-ai-companion-paula-dozsa-tolan-bvu47mw5k"},"datePublished":"2026-09-15T14:52:52.756Z","dateModified":"2026-09-15T14:53:18.123Z","description":"An engineer from Tolan, a voice-first AI companion app, describes the architectural challenges of building real-time voice conversations with LLMs:...","image":"https://i.ytimg.com/vi/xLUQOqjudtA/sddefault.jpg","thumbnailUrl":"https://i.ytimg.com/vi/xLUQOqjudtA/sddefault.jpg","isAccessibleForFree":true,"articleSection":"AI Engineer","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"AI Engineer","logo":"https://media.daily.dev/image/upload/s--u5PucxNT--/f_auto/v1724338940/logos/aidotengineer","url":"https://daily.dev/sources/aidotengineer"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/tolan-voice-first-ai-companion-paula-dozsa-tolan-bvu47mw5k","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai-agents,claude,voice-ai","timeRequired":"PT15M","video":{"@type":"VideoObject","name":"Tolan: Voice-First AI Companion — Paula Dozsa, Tolan","description":"An engineer from Tolan, a voice-first AI companion app, describes the architectural challenges of building real-time voice conversations with LLMs:...","thumbnailUrl":"https://i.ytimg.com/vi/xLUQOqjudtA/sddefault.jpg","uploadDate":"2026-09-15T14:52:52.756Z","duration":"PT15M","url":"https://api.daily.dev/r/bVU47mW5k","embedUrl":"https://www.youtube.com/embed/xLUQOqjudtA"}} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"AI Engineer","item":"https://daily.dev/sources/aidotengineer"},{"@type":"ListItem","position":3,"name":"Tolan: Voice-First AI Companion — Paula Dozsa, Tolan"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/tolan-voice-first-ai-companion-paula-dozsa-tolan-bvu47mw5k#faq","mainEntity":[{"@type":"Question","name":"How does Tolan decide which LLM to use for a given conversation turn?","acceptedAnswer":{"@type":"Answer","text":"Tolan runs a tiered model fleet and routes turns based on emotional stakes rather than cost. A lightweight classifier called the tone router, running on a cheap model, reads the emotional state of each turn; high-stakes moments like a user's first message or crisis-level conversations always get the frontier model, while casual back-and-forth and background tasks run on smaller, faster models that cost roughly five times less per turn. Teams weighing frontier versus small-model routing for cost control can track real-world architecture patterns like this on daily.dev."}},{"@type":"Question","name":"What made the biggest improvement in voice AI response latency for Tolan?","acceptedAnswer":{"@type":"Answer","text":"Switching to GPT-5.1 on the responses API cut time to speech by more than seven-tenths of a second, the single biggest latency improvement reported. This mattered because the full pipeline (end-of-utterance detection, transcription, first token, text-to-speech, playback) needs to complete in under about two seconds total, since a half-second increase in latency measurably hurt user retention metrics. Developers chasing real-time LLM latency budgets can follow model-swap wins like this one on daily.dev."}},{"@type":"Question","name":"How does an AI companion app handle long-term memory without blowing up context length?","acceptedAnswer":{"@type":"Answer","text":"Rather than storing full conversation transcripts, Tolan treats memory as a retrieval system: facts, preferences, and emotional signals are extracted, embedded, and stored in a vector database with sub-50-millisecond lookups. Nightly jobs merge duplicates, cluster related memories, resolve contradictions, and drop noise, while memory is split into a volatile live-tail and a stable, cacheable block based on what actually gets recalled. Anyone designing memory systems for conversational agents can track retrieval-based approaches like this on daily.dev."}}]} ```