How we actually build it
Astrael is an AI company. Below is what we actually build — real measurements, real architecture, real trade-offs.
Astrael-1 + Claude · why we built two engines instead of one
Astrael runs on a dual-engine model layer. The Claude family (Sonnet 4.6 / Opus 4.7 / Haiku 4.5) covers long-form psychological synthesis. Astrael-1 — our own LoRA fine-tune series on leading open-source foundation models — covers Chinese 命理语感, BaZi/Zi Wei nuance, and the small-but-loaded cultural moments where translation-stiff LLM Chinese reads like a tourist. A router decides which engine renders which turn. Both can be active inside a single session.
Self-critique loop · how the Mythic tier sharpens before it streams
Mythic-depth readings go through a three-stage write → critique → revise loop · primary model drafts, Haiku 4.5 reviews, primary rewrites. Roughly 1.7× the cost of a single-pass call · gated to the tiers whose plan headroom absorbs it.
Why Astrael routes between Sonnet 4.6, Opus 4.7 and Haiku 4.5
Different tasks deserve different brains. Routing per turn — instead of always using the top-tier model — lowers latency 4×, cuts cost 8× and keeps response quality where it matters most.
Why we don't let the LLM compute planetary positions
LLMs hallucinate astronomy. Birth charts, BaZi pillars and I Ching casts all run on deterministic libraries server-side. The LLM never sees a coordinate it has to compute.