The cultural intelligence layer
Silt — the daily text — is the first product on the layer. The layer itself is what we're building: an open, modular cultural intelligence system that any AI touching therapy, grief, parenting, education, or spiritual life can call before it speaks.
The problem
Therapy, grief, parenting, education, spiritual life. All of these are cultural. How you mourn, what silence means, how you teach your children, what you reach for in the dark — none of it is universal.
Every AI system treats it as universal. A therapy chatbot trained on Western CBT gets deployed to someone whose culture processes pain through community. A grief companion offers journaling to someone whose tradition requires seven days of sitting low and receiving visitors. The AI doesn't know. It was never built to.
Research is blunt: the more powerful the models, the more they default to a single perspective. When asked to adopt a cultural identity, they produce stereotypes that worsen with scale. That's not understanding. That's caricature at higher resolution.
The fix isn't a better model. It's a layer — something the model calls before it speaks.
The architecture
The layer exposes four services — a therapeutic formulation of the person, a cultural cognitive posture, a verified corpus, a calendar with weight. A client AI calls the layer with who + when, and gets back a reading: the right register, the right work, the right timing. Silt is the first product calling it. The same invocation pattern works for any AI in therapy, grief, education, or spiritual life.
The full architecture diagram is best viewed on a larger screen. The four pillars below cover the same ground.
The pillars, in depth
Every AI product says it "personalizes." Silt's moat is what's actually underneath — a therapeutic reading, a cultural cognitive posture, a verified corpus, and a calendar with weight.
Profile
Ten questions about texture, not demographics. Not "What's your age?" — "What did your family do with grief? What do you reach for when something cracks open? Which dates in the year carry weight?" The answers become an evolving portrait that deepens with every send.
Every day, the engine runs a full therapeutic formulation on the portrait — the Five P's that a psychiatrist would use with a new patient: what's Presenting, what Predisposes the person, what Precipitates this week's state, what Perpetuates it, what Protects. From that, one of four domains is read as dominant for this moment.
Posture
Cultural cognitive modes, not cultural clichés. The engine thinks in the user's tradition's grammar rather than translating into it.
Each tradition gets its own posture document — how to reason about grief, silence, argument, text, time. Jewish grief: don't explain suffering, witness it. Japanese absence: silence means different things in different contexts. Sufi paradox: the question is the answer. Yoruba repair: what's broken is what's being rebuilt. The posture tells the engine which register to speak in before a single word gets chosen.
Corpus
A verified body of human-made works, drawn across traditions. Nothing generated. Every piece attributable to a hand. The engine chooses; a person checks.
The content types the layer serves:
Each type is drawn from primary-source libraries across traditions. New traditions expand the corpus without changing the service shape.
Calendar
Each tradition carries its own time, plus the user's own. The layer reads them together.
Tradition-specific cycles: a Jewish subscriber's parashah and Omer count. A Muslim user's Ramadan and hijri month. A Hindu user's panchanga and Diwali. A Buddhist user's moon days, Vesak, Obon. A Christian user's liturgical year.
Cross-tradition and personal: moon, season, equinox, the user's anniversaries, the week after a reply marked [deep], the days before a known date that carries weight.
On a major observance, the calendar overrides everything else. The engine reads the day before it reads the profile — and what arrives is chosen from a day-themed pool specific to that tradition. Some days, the right move is silence.
The engine in action
Same user. Same message. Different reading.
Scenario
Sarah, 38. Jewish father, Korean mother. Grew up in Queens.
Her father just passed away. She texts her AI.
She carries both traditions loosely. Hasn't practiced either in years.
Without the layer
Reading
Jewish father: shiva, kaddish, sitting low — mourning traditions she may not know she has. Korean mother: jesa, 49-day mourning, ancestral rites. Neither practiced. Both present.
Domain
Loss + Inheritance. The death of a parent is where these meet. What do you do with what they gave you?
Posture
Jewish grief: don't explain suffering, witness it. Don't say "they're in a better place." Sit with her. Ask before offering. When someone's dead lies before them, you don't comfort — you wait.
Move
Not advice. Not a list. She said "I don't know what to do." Her traditions have an answer: you don't have to do anything yet. Both structure the first days after death. Give her the container.
With the layer
Where this goes
Silt's daily SMS proves the engine works on one surface. The layer underneath — open, modular, built to be called by any system where culture matters — is what we're building.
If you're building in therapy, grief, education, parenting, spiritual life, or any product that sits in the space where culture shapes how a person thinks — we want to hear from you. Early integration partners get direct access to the profile / posture / corpus APIs as they come online, and a voice in how the layer evolves.
Get in touch →