The San Mateo company teaching machines to read the feelings people never say out loud - and pointing that skill at healthcare instead of ad-tech.
Here is a thing that is true and slightly uncomfortable: it is very hard to know how someone feels, and we have decided to try to solve this with linear algebra.
IPMD, Inc. is a company that reads emotions. That sentence should make you a little suspicious, because "AI that reads emotions" is the kind of claim that has, historically, ranged from over-promised to actively worrying. But IPMD is an unusually specific version of the idea, and its origin story is unusually specific too. In 2015, on New Year's Eve, a UC Berkeley graduate intern working with the company's founder lost her battle with depression. Nobody had recognized the silent version of what she was carrying. The stated mission of the company that grew out of that loss is, more or less, to build software that would.
That founder is Min Lee, a UC Berkeley alum who is IPMD's Founder, CEO and President, and who says he personally logged more than 30,000 hours training the model - some of it on his own face and his own feelings, which is a detail that is either touching or slightly alarming depending on your mood, and which the software would presumably classify accordingly.
The flagship product is called Project M. It uses custom convolutional-neural-network algorithms to sort a human face into eight primary emotions - anger, contempt, disgust, fear, happy, neutral, sad, and surprise - and then sub-divides each of those into roughly four intensities, which the company multiplies out to about 400,000 distinct emotional variations. The number is large enough to be impressive and specific enough to sound engineered, which is the point.
The part that is genuinely hard, and the part IPMD spends the most words on, is not the neural network. It is the labeling. Nearly 500 people, most of them UC Berkeley students, spent something like 130,000 hours looking at faces and agreeing on what they meant, producing more than 200,000 accurately labeled data points. Emotion is subjective; a model can only be as good as the humans who told it what "contempt" looks like. IPMD's actual moat, in other words, is a decade of people patiently arguing about faces.
The company reports crossing 95% and later 96% ROC/AUC scores, and - in a move that tells you something about its confidence and its budget - it publicly benchmarked Project M against Microsoft's emotion recognition. A ~33-person team in San Mateo lining up against a trillion-dollar cloud provider is the sort of comparison you only publish if you think you look good in it.
Technology should enhance human connection, not replace it.
What separates IPMD from the broader "emotion AI" pack is where it points the camera. The obvious commercial use of reading faces is advertising and surveillance - noticing what makes you flinch so someone can sell to you. IPMD instead aims its technology at healthcare: telemedicine, mental-health support, and remote diagnostics, where knowing that a patient is quietly not okay is the entire job. That is a constraint disguised as a mission, and it shapes the product.
The newer platform, EchoAI, extends Project M from faces alone to a multimodal read - facial expressions, language patterns, tone, and micro-expressions - and returns a personalized emotional-analysis report within seconds. The company describes it as a "clinical data engine" and, more poetically, as a place "where technology, humanity, and dignity meet." Under the EchoBloom initiative, IPMD says it wants to walk EchoAI through the FDA process and turn empathy into a regulated medical device, which is a sentence that should stop you, because regulated emotional AI is a much more serious - and much slower - thing than a flashy demo.
There is a second, more commercial line of business: an emotional-AI interview platform that layers emotion detection over virtual hiring conversations, letting recruiters see the feeling behind an answer and adapt their questions. This is the use case most likely to raise eyebrows about bias and fairness, and IPMD talks about "bias-free" facial analysis, which is a claim the whole field is still trying to earn. It is worth watching closely - both because it is where the near-term revenue is, and because it is where emotional AI is easiest to get wrong.
So what can you actually do with IPMD? If you run a telehealth or mental-health service, you can add a layer that flags emotional states a clinician might miss on a small screen. If you hire at scale, you can surface emotional signal in interviews - with all the responsibility that implies. And if you are building anything where a machine needs to notice that a human is struggling, IPMD is a company that has spent ten years, one very personal loss, and 200,000 labeled faces trying to make that noticing reliable.
One emotion engine, several front doors - healthcare, hiring, and expression.
CNN-based engine that reads eight primary emotions and ~400,000 intensity variations from facial micro and hidden expressions, at reported 95-96% ROC/AUC accuracy.
Human-centered agentic emotional AI that reads face, tone, language and micro-expressions, returning a personalized emotional-analysis report in seconds.
Adds emotional signal to virtual interviews so recruiters can gauge responses and adaptively deepen questioning.
Brings real-time emotional insight to remote diagnostics, aimed at underserved and hard-to-reach care.
Pairs generative art with emotion AI to restore expression for people with disabilities - and the FDA medical-device pathway.
Figures IPMD reports about its own model. Read them as company-stated, not independently audited.
Recognized Google technology / cloud build partner supporting EchoAI infrastructure and R&D.
Member of NVIDIA's program for AI startups.
Building a Medical Agentic AI that fuses EchoAI's emotion detection with medical knowledge.
Named an M7 2022 cohort finalist; founder honored as a visionary CEO in emotional AI.
See the emotion engine run in real time.