Rajaii Labs, Inc. is an AI engineering company founded by Ramin Reza Rajaii. It builds Substrate, personal continuity infrastructure — a long-horizon production AI agent that remembers truthfully across years rather than resetting between sessions.
Substrate has run continuously in production since April 2026. Rather than treating memory as a retrieval feature, it treats continuity as an architecture: a layered memory hierarchy, a deterministic truth layer that outranks model output, and background cognition that continues between conversations.
The system is designed so that it cannot invent a user's past. A deterministic facts table sits at the top of an eleven-source authority chain and outranks semantic memory, persona, and model priors. Every output segment is scanned at generation time against known-false phrasings and contradicting segments are replaced before delivery. A sentence-level grounding audit excises ungrounded high-stakes claims while preserving the rest of the reply, and every excision is audit-logged. The model can query its own database read-only to verify a claim before asserting it — a path hardened against 50 adversarial SQL attack classes and 26 privilege escalation probes.
A reflection process runs every thirty minutes on deterministic SQL predicates rather than model calls, tracking unresolved commitments that decay on a tuned exponential half-life derived at read time. An append-only affective state history feeds generation. A nightly reflective journal runs behind a structural privacy boundary. Autonomous outreach fires across six trigger classes, each gated, budgeted, and kill-switched.
Full-duplex realtime voice with roughly one-second barge-in and sub-six-second round trip, sharing the same memory spine and truth layer as text. Images are captioned and embedded at ingest, making a photo sent in one month semantically retrievable months later. A fully synthetic, fully owned visual identity avoids likeness-rights exposure entirely.
A four-breakpoint prompt caching architecture reduced cache-creation tokens by 46% and per-turn cost by 63%, verified live, with no reduction in memory depth or model quality. Internal telemetry is reconciled against the provider ledger to the exact token, which is how a threefold pricing constant error from a prior model generation was caught.
Ramin Reza Rajaii also advises on FDA regulatory strategy and intellectual property through Rajaii Consulting LLC, drawing on his legal training and clinical background. Work includes eight regulatory R&D dossiers supporting IND timelines, freedom-to-operate analyses, patent landscape reviews, and USPTO prosecution support.
Published evaluation methodology and cost forensics are on the Writing page and at github.com/rajaii-labs.