Asha|AI

How Asha|AI Works

Asha doesn't generate answers from a language model's memory. It retrieves evidence from a structured knowledge base of 126M+ verified vectors across 829 indexed collections, then synthesizes a response with full citations.

The Evidence Pipeline

When you ask Asha a health question, here's what happens behind the scenes:

Step 1: Understand Your Question

Asha parses your question to identify the medical concepts, entities, and intent. It works in many languages via Gemini 3 multilingual, automatically detecting the language and preserving clinical meaning across translations.

Step 2: Search the Knowledge Base

Your question is converted into a semantic vector and searched against 126M+ Cognitive Inference Units (CIUs). These are structured knowledge atoms drawn from PubMed, OpenAlex, PMC, StatPearls, DailyMed, FDA, and clinical guidelines. Each source is weighted by clinical relevance: StatPearls 1.5x, DailyMed 1.5x, clinical guidelines 1.4x (and 1.5x for the v3 stack).

Step 3: Validate and Cross-Reference

No single knowledge atom can vouch for itself. Every recommendation requires validation from a second, independent source, the way a surgical team confirms it's safe to proceed. Confidence scores and falsifiability checks determine whether a claim is trustworthy enough to include.

Step 4: Synthesize the Response

The validated evidence is synthesized into a clear, plain-language response. Every factual claim links back to its source. If Asha can't find confident evidence, it says “I don't know” instead of guessing.

Step 5: Safety Checks

Before any response reaches you, it passes through safety filters: emergency detection, prescription guardrails, and citation verification. If Asha detects a potential medical emergency, it interrupts with crisis guidance and emergency contact information.

The Knowledge Base

Asha's knowledge comes from structured, verified medical sources. None of it comes from scraping the internet or training on Reddit posts.

48M+
Peer-reviewed papers (PubMed, OpenAlex, PMC)
126M+
Evidence Vectors
829
Indexed Collections

Key Sources

Cognitive Inference Units (CIUs)

The fundamental building block of Asha's knowledge base is the Cognitive Inference Unit (CIU), a patent-pending data structure that stores medical knowledge as structured atoms:

Each CIU has three components:
Name: what it is (the concept or fact)
Form: how it manifests (the evidence and context)
Dharma: what it does and when to use it (the clinical application)

Every CIU carries a cryptographic hash that chains it to its parent, a semantic vector embedding for intelligent retrieval, and a confidence-falsifiability score. Modify any unit and every descendant hash breaks, making the knowledge base tamper-evident by design.

Fiduciary Architecture

Most AI systems are “aligned” through training to be helpful. Asha is fiduciary: architecturally constrained to act in your interest. The difference:

Aligned AI

Trained to be helpful through RLHF and safety filters. Policies can be overridden. The model decides what's safe.

Fiduciary AI (Asha)

Architecturally constrained. Harmful paths are unreachable, beyond discouragement. The architecture decides what's safe, not the model.

This is a cognitive architecture covered by an allowed United States patent (US 19/290,471, Notice of Allowance October 2025) and additional pending applications.

Privacy Architecture

Asha's privacy is architectural, beyond policy:

See It in Action

Ask Asha a health question and see the evidence for yourself.

Try Asha Free