SATI

AI-assisted pattern recognition for jhana meditation practitioners.

The Problem

Jhana practitioners often struggle with the same obstacles repeatedly without recognizing the pattern. A practitioner might describe "trying harder and harder but getting nowhere" without realizing this is the problem—over-efforting blocks the very state they're seeking.

Human teachers can spot these patterns, but teacher time is limited and expensive. Most practitioners get a few hours of guidance per month at best, leaving long stretches where unhelpful patterns go unnoticed.

The Solution

SATI provides pattern recognition between teacher sessions. Practitioners describe their experience in natural language, and the system:

1. Identifies likely antipatterns from a taxonomy of 20+ documented obstacles
2. Calibrates confidence based on how detectable each pattern is from text
3. Adapts response strategy to uncertainty level—advising when confident, clarifying when not
4. Tracks patterns across sessions to surface what's recurring

Current Performance

These numbers reflect honest limitations. The system gets the top pattern right about two-thirds of the time, which is why we surface multiple hypotheses and hedge appropriately when uncertain. See Evidence for full metrics and limitations.

How It Works

1User Input
2    │
3    ▼
4┌─────────────────────────────────┐
5│  Safety Check (Pattern Match)   │ ──▶ Escalate if crisis detected
6└─────────────────────────────────┘
7    │
8    ▼
9┌─────────────────────────────────┐
10│  LLM Classification             │
11│  ├─ Structured output (Zod)     │
12│  ├─ Reasoning traces            │
13│  └─ Multiple hypotheses         │
14└─────────────────────────────────┘
15    │
16    ▼
17┌─────────────────────────────────┐
18│  Confidence Calibration         │
19│  ├─ Per-antipattern multipliers │
20│  └─ Bucket assignment           │
21└─────────────────────────────────┘
22    │
23    ▼
24┌─────────────────────────────────┐
25│  Response Strategy Selection    │
26│  └─ advise | suggest | clarify  │
27└─────────────────────────────────┘
28    │
29    ▼
30Streaming Response

The three-stage pipeline ensures safety comes first, classification produces multiple hypotheses with reasoning, and response strategy matches confidence level.

Explore Further

Antipattern Taxonomy

The system recognizes 20+ documented obstacles organized into categories:

Detection difficulty determines confidence calibration—we're appropriately less confident about patterns that are hard to infer from text.