Technology
Three layers, one constraint: the user’s inner life stays the user’s.
Our systems combine modern language models with a personal context architecture developed in-house. The design brief is unusual for consumer software: maximise understanding of one person, minimise what anyone else — including us — can see.
01
Reflective dialogue
The conversational layer is tuned for a specific job: asking better questions. Rather than producing generic answers, the system is evaluated on whether its prompts help users articulate reasoning they had not yet put into words. We maintain internal evaluation suites for question quality, tone discipline, and the boundary between reflection and advice.
02
Personal context engine
Underneath the dialogue sits a long-lived, per-person context store: a structured representation of what the user has shared over time — themes, stated values, recurring situations. Retrieval is scoped strictly to the individual it describes. Nothing in this layer is aggregated across users or used to train foundation models.
03
Privacy and control plane
Every capability above rests on infrastructure designed around data minimisation: encryption in transit and at rest, per-user isolation of reflective context, full export in machine-readable form, and deletion that actually deletes — including derived representations, within a fixed and audited window.
Models
Top models, made ours
We are not a foundation-model laboratory, and we do not pretend to be. Our craft is applied: taking the best models in the world and shaping them — through distillation, fine-tuning, and rigorous evaluation — into something that serves one person deeply and privately.
01
Frontier models, selectively
We build on the strongest available foundation models and are deliberately unsentimental about which — every release cycle, candidates are benchmarked against our internal evaluation suites for reflective dialogue, and only the winners ship.
02
Distillation and fine-tuning
Where a frontier model is more than the task requires, we distil its capability into smaller, faster models under our control, and fine-tune them on carefully curated reflective-dialogue data. The result is lower latency, lower cost, and a tighter privacy boundary around every conversation.
Our own architecture
Agent memory built on semantic cores
The part we build entirely in-house is the memory. Each person’s agent maintains a set of semantic cores — compact, structured representations of the themes, values, and recurring patterns in what that person has shared over time.
Cores ground every retrieval, keep the agent coherent across months of conversation, and — because they are explicit rather than buried in model weights — can be inspected, corrected, and erased by the person they describe.
In development
One personal AI, on every screen you live with
We are building our products as first-class experiences for the web, Android, and iOS — one system, one memory, one privacy architecture, available wherever a quiet moment finds you.
Our market is the global consumer — not enterprises, not specialists. We believe advanced AI belongs in everyday hands, used for something genuinely good: understanding yourself better. That reach is exactly why data security and regulatory compliance sit at the top of our priorities, not the bottom.
- WebThe full experience in the browser — nothing to install, everything encrypted in transit.
- AndroidA native app built for the moments reflection actually happens: on the move, in between things.
- iOSThe same personal AI, with platform-level privacy protections used to their full extent.
Data stewardship
Built as if the data were our own
Reflection produces some of the most sensitive data software can hold. Our handling of it is intentionally boring: strict isolation, no advertising use, no sale of personal data, and no training of underlying models on user content.
Read the Privacy Policy
Engineering practices
The unglamorous work that makes reflective AI trustworthy
01
Evaluation before release
Changes to reflective behaviour pass structured evaluation — internal red-teaming for tone, boundary, and safety regressions — before reaching users.
02
Human review of the system, not the user
Quality assurance is performed on the system's behaviour using consenting test accounts and synthetic scenarios, not by browsing real users' private reflections.
03
Conservative model use
Where we rely on third-party foundation models, we contractually and technically exclude our users' content from provider training pipelines.
04
GDPR as an engineering spec
Data-subject rights — access, portability, erasure, objection — are implemented as product features with service-level targets, not as manual back-office processes.
Technical questions?
Our engineering team reads everything sent to info@antipadia.com.