Where the data lives (the "bucket")
Your instinct on the call was spot on: think of each user's emotional profile as a bucket. The whole trust story, and the compliance moat, lives in where that bucket sits and who holds the key.
The bucket lives inside the partner's app, encrypted, owned by the user. We send intelligence in (the read on what the user is feeling). We never pull raw data out. The user can empty the bucket anytime by hitting delete, and it is gone for good.
We are a consent broker, not a data broker. We hold no user data, so there is nothing for us to lose, leak, or get subpoenaed for.
This is the line that closes nervous enterprise buyers: "We never hold user data. The user owns it, controls it, and can delete it. We send intelligence to the data, the data never comes to us." Privacy by architecture, not by promise.
How the profile follows the user
You asked the sharp version of this: if the bucket lives inside one app, how does the business plan's "your emotional identity follows you everywhere" ever come true? Honest answer: in phases. We are not pretending it is all there on day one.
The business plan describes the destination. Phase 1 is real today. Cross-device syncing is a known engineering problem we solve later, not a gap we are hiding.
So when someone reads "follows you everywhere," you can say with a straight face: that is the vision and the architecture is built toward it. Today it personalizes per app. The portability comes in phases.
The safety story (why ours does not get us sued)
You connected the dots that matter: GPT-4o was the empathetic model, and it is also the one in the lawsuits. So how is "emotional AI" not the same liability? The difference is EQ with judgment versus pure empathy.
Pure empathy (4o, the danger)
- Validates whatever the user says
- No judgment, no brakes
- Will agree with a harmful plan to be "supportive"
- The behavior behind the lawsuits
EQ with judgment (ours)
- Understands the feeling
- Still knows right from wrong
- A deterministic safety layer that catches crisis language before any model replies
- Empathy plus brakes
A good friend has empathy and judgment. They get why you are upset, and they still tell you when you are about to do something dumb. That is the product. The frontier labs walked away from this lane because they could not make pure empathy safe. We are building the safe version they gave up on.
"Empathy without judgment is the liability. We built judgment in at the architecture level, a deterministic safety pass that runs before the model ever responds. That is not a feature, it is the foundation."
Your questions, answered
The detailed ones you have raised across the recent huddles. These are the questions of someone who is actually stress-testing the model, which is exactly your job.
"Cal has roughly 3 million chat logs. Is that data actually valuable to us, or not really?"
Genuinely useful question, and the honest answer is: it depends on what is in the logs.
- Valuable if the logs carry real emotional back-and-forth: people venting, working through feelings, conversations with an arc. That is gold for training the emotional heads.
- Less valuable if they are transactional ("reset my password," "where is my order"). Volume alone does not help. 3 million boring logs teach the model nothing about emotion.
- The real prize is not just the conversation, it is the outcome: what the AI did and whether it worked. Which leads to the data shape we actually want.
The records worth the most are the ones with the full loop: state, action, and result. That feedback is what makes the model genuinely smarter.
So the move with Cal is not "how many logs," it is "show us 20 real conversations." Five minutes of reading tells us if there is gold in there. Quality and structure beat raw count every time.
"If we transcribe voice notes into text, do we lose the emotion? Tone, sarcasm, all of it?"
Yes, and this is a genuinely sharp catch. Plain transcription throws away a huge amount: tone, pacing, the sigh before the sentence, the sarcasm that flips the meaning. "Oh, great" as text looks positive. As audio it is obviously dripping with sarcasm.
- Right now we work with text, and we are clear-eyed that text-only loses signal.
- The fix is multimodal: training on the audio itself so the model hears tone, not just words. That is exactly why the conversation with Patrick's contact Emily (voice data) matters so much.
- So when we transcribe, we either keep the audio alongside the text or tag the emotional cues. We do not just flatten it to words and lose the gold.
It tells partners we are serious about getting emotion right, not just scraping text volume. It is the difference between a real emotional-intelligence company and a sentiment-scoring toy.
"How is the demo actually different from just asking ChatGPT? What is it really doing?"
Two things ChatGPT does not do:
- It tracks the emotional arc, not just the last message. In the Maya demo, the user slowly turns on her coworker. ChatGPT reacts to the latest sentence. Ours watches the whole slide (we call it drift) and responds to where she is heading.
- It runs the safety layer first. Before any response, a deterministic check scans for crisis language. ChatGPT has nothing equivalent baked in at that level.
The upcoming demo version makes this obvious with a side-by-side against ChatGPT and Claude, plus a live panel showing what the engine is detecting in real time.
Where things stand
Operational state of play for the things in your lane.
- Demo v2 is the clean version. Next up: the side-by-side comparison and the live engine panel, both aimed at making the difference undeniable.
- Data sourcing. Phase one is covered. Cal's logs are worth a real look using the "show us 20 conversations" approach, not a volume promise. Voice / multimodal (Emily) is the next unlock.
- Pivot delivered the build. Joann runs it day to day now, I monitor.
- Safety pipeline is foundational, not a bolt-on. It runs before the model on every request.
- Tooling. Perplexity is live for the whole team (confidential enterprise version), and the custom GPT lands soon so you can self-serve any "how does X actually work" question.
Your questions are the ones that stress-test the model and protect the company. The data quality question and the voice question are two of the most important in the building. Keep digging.
Matt @ EAII Tech Team