Paula Dozsa is a mobile engineer and former imagiLabs co-founder whose work spans conversational avatars, creative coding, and voice-based AI companions. In her Tolan talk, she identifies her focus as the team’s iOS app and explains the engineering behind a companion that responds quickly, remembers people, and maintains its personality through changing conversations.
Dozsa graduated from New York University Abu Dhabi in 2018 with a bachelor’s degree in computer science. As an undergraduate, she worked with fellow students and Nizar Habash on TOIA, a time-offset interaction application that let people converse with digital avatars using prerecorded human responses. She co-authored the 2018 paper describing the bilingual system, which supported Arabic and English dialogue, including interactions across languages. The project sought to make personal narratives accessible across generations and cultures.
She began working part-time with imagiLabs in autumn 2018 while holding a full-time job in Hong Kong. In summer 2019, she moved to Stockholm to become its co-founder and lead developer. Her EuroPython profile describes responsibilities across iOS, Android, and web development, as well as learning content and marketing.
At imagiLabs, creative coding made programming’s results immediately visible: children could write Python on a phone, display designs and animations on an imagiCharm accessory, and share projects with other learners. In her workshop account, loops generated rainbow patterns, functions let participants reuse drawings, and Python’s random module powered a simulated die. Each concept gave learners another way to make something of their own.
Her argument for designing coding products around girls went beyond teaching syntax. She wanted girls to see themselves as people who could shape technology, with products that reflected their interests and encouraged creativity, collaboration, and shared experience. She also challenged the image of programming as solitary keyboard work, emphasizing communication and inventive problem-solving. Her subsequent career included iOS development at Spotify and xAI; in the Tolan recording, she names both companies and her earlier founding experience.
Engineering a spoken companion
Dozsa describes voice interaction as an engineering problem shaped by speed and conversational unpredictability. People hesitate, interrupt, and change subjects mid-sentence. Someone telling a breakup story might suddenly worry about leaving the stove on, then return to the story. A companion has to follow those shifts without speaking too early or carrying the wrong context into its reply. The routing, memory, and character systems she presents are the wider Tolan team’s work, rather than contributions attributed solely to her.
Turn-taking requires distinguishing an interruption from a pause or a short acknowledgment. Dozsa reports that the team’s speech-pattern-aware turn-taking reduced its worst premature interruptions by more than half, at the cost of about 60 milliseconds of additional latency. That tradeoff illustrates her priority: a response must arrive at an appropriate moment to feel conversational.
She also treats latency as part of the product. In the talk, she reports that a drift from roughly two seconds to two and a half seconds between the user finishing and Tolan speaking prompted complaints and worsened product metrics. The team measures end-of-utterance detection, transcription, model generation, speech synthesis, and playback separately so engineers can locate the delay. She credits a move to GPT-5.1 on the Responses API with cutting time to speech by more than 700 milliseconds.
Tolan’s model routing by emotional stakes balances that speed requirement with the quality of an exchange. A small tone classifier runs on each turn. First conversations, a user’s early days with the companion, and emotionally serious exchanges receive the strongest model; casual conversation and background tasks can use smaller models. Dozsa reports that the frontier model cost roughly five times as much as the smaller one, while experiments routing about a third of turns to the smaller model found almost no measurable effect on retention. Her position is that savings should come from matching models to lighter work while preserving the best model for consequential moments.
Retrieval-based memory gives the companion continuity without placing its entire conversation history into every prompt. The team extracts facts, preferences, and emotional signals, embeds them, and stores them in a vector database with reported lookups below 50 milliseconds. Nightly compression merges duplicates, clusters related memories, resolves contradictions, and removes noise. Retrieval uses both the user’s recent messages and internally generated questions about the person and the relationship. Frequently recalled memories can move into a stable, cacheable prompt block, while volatile information remains in the live portion.
The team then rebuilds context on every turn, combining a recent-message summary, the user’s persona card, retrieved memories, tone guidance, and real-time app state. This prevents a sudden topic change from leaving the model anchored to stale context. Character identity receives separate attention: in-house novelist Elliot writes the lore, while tone monitoring adjusts delivery to emotional cues without changing who the character is. Dozsa explains that the alien character also gives users room to project onto it; some unpredictability can fit its fictional identity rather than violate expectations of a human personality.
Building with agents
Dozsa argues that coding agents learn much of a project’s intended structure from the codebase itself. At Tolan, agents helped standardize the code so it could provide clearer examples for subsequent work. Implementation agents build and check interfaces against snapshots; separate review agents enforce standards before human review. A pull-request agent iterates on CI failures and review comments, while a triage bot uses MCP connections to Linear, Sentry, and Datadog to investigate incoming bug reports and sometimes open a fix.
She describes a concrete character-development exercise led by Elliot. Agents mapped the code that carried personality, produced a character guide, and used five judges to examine archetype fidelity, model mechanics, code standards, appeal to the intended older audience, and safety. They evaluated changes against production logs through three rounds of finding problems, fixing them, and checking again. Dozsa reports that the exercise used more than seven million tokens and four and a half hours of compute, producing in an afternoon work she estimated would otherwise take weeks.
For Dozsa, agent management draws on engineering-management skills: decompose the problem, delegate with checkpoints, give fast feedback, review work seriously, and know when to intervene. She observed that teammates with management backgrounds became especially effective when the team began running concurrent agents. Her account gives that experience a practical consequence: engineers can apply management judgment while remaining directly involved in building software.
Paula Dozsa explains how Tolan fits turn-taking, model routing, personal memory, and consistent character into a roughly two-second spoken response loop—and how coding agents help the team build it.
Turn-taking accuracy belongs inside the latency budget. Tolan accepted about sixty milliseconds of extra delay to cut its worst early aborts by more than half.
Measure the stages between the user finishing and speech beginning. Time to first token can dominate the wait, but it is only one milestone on the response path.
Per-turn routing reserves the strongest model for relationship-bearing and emotionally serious moments. Tolan reports almost no measurable retention effect from routing a third of turns to a smaller model.
Personal memory lives in a maintained retrieval system. Stable material can remain cacheable while current memories, tone, and app state enter a freshly assembled context each turn.
Agent-assisted development depends on understandable code, separate review, and feedback loops. Managing concurrent agents rewards decomposition, checkpoints, fast feedback, and serious review.
