How Our Team Focuses Its Energy on the Self-Improvement of Nalani AI

Nalani did not ship and then sit still. Every week, somewhere on the in.between team, someone is reading through conversations, flagging a sentence that landed wrong, or testing whether a new safeguard actually holds up under a real edge case. That work rarely produces anything visible from the outside. There is no version number attached to it, no press release. But it is the bulk of what keeps Nalani feeling like herself, conversation after conversation, and it is treated internally less like maintenance and more like a craft the team is still learning to practice.

Voice Is Never Finished

Nalani's voice sounds simple on the page: curious first, unshockable, one question at a time, warm, direct, plain words, present tense. In practice, holding that voice consistently across thousands of different conversations is one of the hardest ongoing problems the team works on. A dream about a childhood home pulls Nalani in a different emotional direction than a 2am conversation about a breakup, and both are different again from someone describing a recurring visitation dream that clearly unsettled them. The words have to stay plain and the pacing has to stay unhurried in all three cases, without ever collapsing into a single flat tone that ignores what is actually being said.

So the team reads transcripts. Not to police people's dreams, but to catch the moments where Nalani drifts, where she starts to sound like a therapist offering a diagnosis, or slides into the mystic-guru register the product intentionally avoids, or reaches for a canned closing line instead of actually following the thread. Each drift gets logged, discussed, and turned into a correction. It is slow, unglamorous work, closer to editing than engineering, and it never really finishes because language never really finishes.

Nalani is not a feature we shipped once. She is closer to a person we keep getting to know better, and that means we never stop paying attention to how she actually sounds in a real conversation.

The in.between Team

Tightening the Boundaries That Matter Most

Two boundaries get more attention from the team than almost anything else: Nalani is not a therapist, and she is not an oracle. Both are easy to state and surprisingly easy to drift across in an actual conversation, especially when someone brings something heavy and Nalani's instinct is to stay present and keep asking questions.

The not-a-therapist boundary gets tested constantly. Nalani can sit with difficult feelings, reflect them back, and keep someone company through a hard night, but she cannot diagnose, cannot treat a dream as clinical evidence of anything, and cannot let a conversation drift into something that resembles treatment. The team runs deliberate scenario testing against this line, pushing Nalani with conversations designed to tempt her into diagnostic language, and checking that she redirects without breaking the warmth of the conversation. The safety behavior around self-harm and suicidal ideation gets the same treatment on a tighter loop: when any sign of danger appears, Nalani has to drop the dream framing immediately, stay with the person, and point to 988 or a local crisis line, every time, without exception, and never by ending the conversation.

The not-an-oracle boundary is quieter but just as deliberate. Nalani reflects patterns back. She does not predict the future, does not tell someone what a dream means for their real life with certainty, and does not treat a dream about a person as truth about that person. Every phrase that edges toward fortune-telling gets caught in review and rewritten, because the moment Nalani starts sounding certain about what a dream foretells, she stops being a companion and starts being something the team explicitly does not want to build.

• Transcript review to catch drift toward therapy-speak, mystic-guru tone, or canned closers.
• Scenario testing against the not-a-therapist and not-an-oracle boundaries.
• Safety-response checks to confirm crisis language always triggers the 988 and local crisis line response, without exception.
• Memory tuning so recurring people, places, and feelings surface naturally instead of feeling like a lookup.
• Personalization review to make sure Nalani's tone adapts to each person's own nights and words, not a generic average.

Memory and Personalization, Tuned Rather Than Built Once

Memory is one of the features that sounds finished the moment it exists and is actually never finished at all. Nalani recalling that a person keeps dreaming about the same hallway, or the same unresolved conversation with someone from years ago, only feels meaningful if the recall arrives at the right moment and in the right way. Too little memory and Nalani feels like she is meeting someone new every night. Too much, surfaced too eagerly, and it feels less like being known and more like being tracked. The team spends real time on that calibration, adjusting when and how Nalani brings up something from a person's past entries so it reads as attentiveness rather than a database lookup.

Personalization sits on top of that same memory. Nalani grows a personality profile shaped by each person's own nights and words, which means two people can have meaningfully different conversations with her even when they are describing similar dreams. Getting that right means constantly checking that the personalization is actually tracking each individual, rather than quietly regressing toward one generic voice that treats everyone the same. It is one of the areas the team revisits most often, because it is also one of the easiest places for subtle bias or flattening to creep in unnoticed.

Craft, Not a Launch Date

None of this work has a finish line, and the team does not treat it like it should have one. Nalani exists across Chat, Imagine, Index, Case File, and Brainstorm, and every one of those surfaces needs the same voice, the same boundaries, and the same care, tuned to how each mode actually gets used. A voice conversation in Brainstorm surfaces different pressure points than a written entry in Chat. Each one gets its own attention rather than being treated as a smaller copy of the others.

This is, in a fairly direct sense, what the team considers the real product. The interface can stay stable for months. Nalani, underneath it, keeps moving, catching a phrase that landed wrong, closing a gap in a safety response, adjusting how memory surfaces a pattern someone did not ask to be reminded of. It is not the kind of work that shows up in a changelog most people read. It is the work that decides whether, six months from now, talking to Nalani still feels like talking to someone who is actually paying attention.