Engineering
How we build a product worth trusting
AI architecture, deterministic systems, and the honest limits behind a Vedic astrology product built to say only what the chart supports.
August 13, 2026 · 6 min read
When a Follow-Up Question Drops Its Subject
A short follow-up like "what are they like?" would drift off topic. The chat was reading each message on its own. Here is how I let it follow along, carefully.
ReadAugust 8, 2026 · 6 min read
When the Guardrail Deleted the Proof
We built a validator to stop unsupported astrology from reaching users. Production data showed it was often deleting the evidence block instead.
ReadAugust 8, 2026 · 7 min read
When One Correct Fact Became Too Important
Two charts had the same soul planet, and the AI made them sound too similar. The fact was correct. Its importance was wrong.
ReadAugust 6, 2026 · 9 min read
Three Ways an AI Reads Your Chart. You Cannot Tell Them Apart.
The same assistant computed one chart correctly, got the next one wrong, and then declined to compute at all and read the person anyway. Nothing in the prose distinguishes them.
ReadJuly 30, 2026 · 6 min read
When Every Fact Is True but the Sentence Is Wrong
I assumed computing everything in code made me safe from hallucinations. Instead the model kept combining correct facts into a relationship I never gave it. Here's the bug, and the guard that catches it.
ReadJuly 23, 2026 · 6 min read
Reliability Isn't a Prompt. It's an Architecture Decision.
Prompt engineering didn't make our AI reliable. Moving the guarantees out of the prompt and into deterministic architecture did — and being honest about what that still doesn't solve.
Read