Solving Voice Onboarding

Imagine signing up for a dating app, and instead of answering twenty questions about your hobbies and deal-breakers, you just... talk. For a few minutes. And the AI on the other end already knows you read the Daily Stoic every morning, that you'd rather be in a quiet coffee shop than a packed bar, and that your idea of a perfect evening involves Vinland Saga and a deep conversation about philosophy. It doesn't ask you these things. It already knows.

That's what we built.

We took everything Onairos knows about a user and fed it directly into the onboarding agent's brain before the conversation even starts. Personality traits, compressed memories from across their apps, and real-time inference from a model trained specifically on their behavior. The result? The agent reaches matchmaking readiness in roughly half the time, asks ~60% fewer questions, and the profiles it creates are significantly richer than anything a cold-start questionnaire could produce.

This isn't a small improvement. It's a fundamentally different way for AI agents to learn about people.

The Problem Everyone Ignores

Most dating apps still follow the same tired playbook:

1. User signs up. Blank profile.
2. The app (or an AI agent) starts asking questions. Hobbies? Values? Deal-breakers? Ideal first date?
3. User answers five, ten, maybe twenty times. Usually by typing.
4. Only after all that does the system have enough information to suggest anyone remotely compatible.

The numbers tell the story: 40 to 60% of users drop off during signup flows. The ones who stick around give surface-level answers because, honestly, who wants to write an essay about themselves to an app? And even the newer voice-enabled versions make the same fundamental mistake. They treat every user as a blank page, asking redundant questions even when the signals already exist somewhere.

The real issue is context poverty. The agent starts the conversation knowing absolutely nothing, so it burns through turns confirming basics instead of exploring what actually matters for compatibility.

Why Voice Changes Everything

Voice is a better interface for this, and it's not even close.

When someone talks about their love for stargazing, you can hear whether it's genuine passion or just a polite answer. Tone, enthusiasm, hesitation, the way someone's voice lifts when they mention something they truly care about. Text strips all of that away. And practically speaking, people are more willing to talk than type. Voice completion rates run 20 to 30% higher in most apps because you can do it while walking, cooking, or lying on your couch.

But voice alone doesn't solve the underlying problem. If the agent still doesn't know anything about you going in, it's just asking the same twenty questions out loud instead of on screen. You need voice plus context.

How We Actually Solved It

We give the agent three layers of insight about you before it says a single word:

All of this gets injected directly into the system prompt. The agent goes from being a "blank slate questioner" to an informed conversational partner before hello.

Prompt Augmentation in Action

You're onboarding a user for dating matches. Have a genuine conversation
to understand them. Ask whatever feels relevant based on what they share.
Trust your judgment on what matters.
Once you genuinely feel you know enough to match them well, wrap up.

Critical Instruction:
Always check context before asking. Complete onboarding ASAP.

[Same base prompt]

Personality Traits of User:
{"Stoic Wisdom Interest": 80, "Daily Stoic Engagement": 90,
 "Philosophical Discussions": 85, "AI and ML Enthusiasm": 40...}

Memory of User:
Reads Daily Stoic every morning... Prefers small coffee shop meetups...

MBTI (Personalities User Likes):
INFJ: 0.627, INTJ: 0.585, ENFJ: 0.580...

Critical Instruction:
Always check context before asking. Complete onboarding ASAP.

Both agents get the same base prompt and the same closing instructions. The default agent simply has no context to work with, so it has to ask everything from scratch. The Onairos agent skips redundant questions, respects what it already knows, and focuses only on gaps—reaching saturation in significantly fewer turns despite receiving no special behavioral advantage in the prompt itself.

Not Just Faster—Fundamentally Better Matchmaking

Speed is only half the story. Because the agent saturates with deeper, pre-enriched context:

• Nuance & Precision — Traits + memories capture subtleties (e.g., "turned off by superficiality" + "drawn to intellectual depth") that pure question-asking often misses.
• Reduced Bias & Better Generalization — Inference models spot patterns across users; compressed memories preserve signal without noise.
• Post-Saturation Quality — Matches improve because the profile reflects authentic preferences, not just answered questions.

Faster onboarding + richer profiles = dramatically better long-term outcomes.

Results

We ran side-by-side voice onboarding sessions using OpenAI's Realtime API via WebRTC. Critically, both agents receive the exact same base system prompt and the exact same critical instructions—including the directive to avoid redundant questions and complete onboarding as efficiently as possible. The playing field is level. The only difference: the Onairos panel also receives pre-injected personality traits, compressed memories, and MBTI inference scores. Every improvement you see below comes purely from the context advantage, not from prompt engineering tricks.

Breaking Down the Efficiency Gains

The average onboarding session with the default agent ran 2 minutes and 20 seconds, with some sessions stretching past 3 minutes. With Onairos context pre-loaded, that dropped to an average of 1 minute and 32 seconds, with the best runs completing in around 1:30—a reduction of up to 50%.

On average, that's a 48-second reduction per session. In the best observed cases (3:00+ down to ~1:30), the time savings hit the full 50% mark.

At scale, these savings compound dramatically:

• 1,000 daily signups = 800+ minutes (~13 hours) of cumulative user time saved per day
• 100,000 users/month = 80,000+ minutes (~1,333 hours) of reduced onboarding friction
• Each saved second directly reduces the window where users abandon the flow

The ~60% reduction in questions asked is equally significant. The default agent typically needed 10–20+ questions to build a sufficient profile. The Onairos-augmented agent achieved equal or better profile coverage in just 4–8 questions, because it already knew the fundamentals and only needed to fill genuine gaps.

These two metrics—time and question count—are the ones we directly measured in controlled side-by-side testing. Profile depth and match satisfaction improvements are expected downstream effects that we plan to validate as the system moves into broader testing.

Under the Hood: The MBTI Inference Engine

The most technically novel component of the Onairos pipeline is the MBTI inference engine. We train a dedicated 214K-parameter model on each individual user's data. Not a fine-tune of a foundation model, but a lightweight, per-user model that learns specifically what that person likes and dislikes from their cross-app behavior.

Once trained, we run inference against 16 MBTI personality type profiles. Each profile contains a description of typical hobbies, activities, and behavioral patterns associated with that type. The model outputs a preference score between 0 and 1 for each, indicating how strongly the user's actual behavior aligns with each archetype's characteristic interests.

Example Inference Output

The agent uses these ranked scores to calibrate conversation style, topic selection, and partner weighting. When it sees high INFJ and INTJ affinity, it knows to explore intellectual depth, philosophical values, and introspective hobbies. It doesn't waste a turn asking "So what are your hobbies?" because it already has a probabilistic map of who this person is.

The Full Context Stack

When you put it all together, the Onairos-augmented agent walks into every conversation with three layers of understanding:

All three layers are injected into the system prompt before the conversation begins. The agent receives the base onboarding instructions, then personality traits, then memories, then MBTI scores, and finally a set of last instructions telling it to always check what it already knows before asking anything and to complete onboarding as efficiently as possible.

Why This Matters for Apps and Users Alike

For Platforms

• Faster onboarding leads directly to lower abandonment, higher activation, and more engaged users
• Up to 50% reduction in onboarding time with 60% fewer questions translates to measurably lower drop-off rates
• Scalable context compression keeps compute efficient even at millions of users
• In a crowded market, intelligent voice onboarding becomes a defensible moat

For Users

• Less effort, more authenticity, fewer bad dates
• You speak naturally; the AI already understands your world
• Quicker path to people who actually get you

The Bottom Line

The best conversations happen when someone already gets you. That's the bar we set for onboarding. Feed the agent everything you've already told us across your apps, and suddenly all those hours users waste answering the same questions over and over again just... disappear. The conversation picks up right where it matters — who you actually are and what you're looking for.

Onairos exists to make your data work for you — better connections, truly personal experiences, and AI that respects who you are.