Sasha AI Engine

Executive Summary

Sasha is ReGenesis's AI engine — an "invisible intelligence" that observes coaching sessions, generates evidence-backed insights, and provides 24/7 companion support to coachees. Sasha is not a chatbot. It is a sophisticated AI system built on Anthropic's Claude that operates within strict safety guardrails, with every output subject to human oversight.

Key Design Principles

Principle	Implementation
Human-in-the-loop	All AI-generated insights require coach approval before client delivery
Evidence-backed	Every claim links to source material (L0/L1/L2 Evidence Packs)
Privacy-first	Data is pseudonymized before LLM calls; no-training flags set on every request
Safety-first	Escalation protocols for crisis, therapy-territory content, and harmful outputs
Transparent	Full provenance logging: model version, prompt template, token counts, timestamps
Killable	Emergency off-switch via feature flag — can disable Sasha instantly per tenant or globally

LLM Provider Strategy

• Primary: Anthropic Claude (Claude 3.5 Sonnet for analysis, Claude 3.5 Haiku for lower-cost tasks)
• Fallback: OpenAI GPT-4o (for provider outage scenarios)
• No model training: Both providers contractually prohibited from training on our data
• Provider diversity: Reduces single-vendor dependency risk

Cost Profile

Task	Model	Estimated Cost Per Call
Session insight generation	Claude 3.5 Sonnet	$0.05 - $0.15
Evidence reasoning (L1)	Claude 3.5 Sonnet	$0.03 - $0.08
Companion chat response	Claude 3.5 Haiku	$0.001 - $0.005
Embedding generation	text-embedding-3-small	$0.0001 per 1K tokens

Working Knowledge

What Sasha Does

Sasha operates in three modes, each requiring escalating consent from the user:

Mode 1: Observe

• Listens to / receives coaching session transcripts
• Stores data but does not process it with AI
• Requires basic session recording consent
• Zero AI cost in this mode

Mode 2: Analyze

• Processes session data through the LLM
• Generates insights, patterns, themes, action items
• Builds Evidence Packs (L0/L1/L2)
• All outputs go to the coach for approval first
• Requires sasha_analyze consent

Mode 3: Act

• Can take real-world actions: schedule meetings, send emails, create tasks
• Every action requires explicit user confirmation before execution
• Highest risk, most restricted
• Requires sasha_act consent + per-action approval

How Sasha Generates Insights

1. Context Assembly: Gathers the current session transcript, previous insights, relevant historical context from the vector DB, and the coaching relationship profile
2. Pseudonymization: Replaces all personal names and PII with tokens before sending to the LLM
3. LLM Call: Sends the assembled prompt to Anthropic Claude with no-training flags
4. Output Validation: Checks the response for safety guardrails, format compliance, and content appropriateness
5. Evidence Assembly: Structures the output into L0/L1/L2 Evidence Packs
6. Cross-Validation: Verifies that claims are supported by source material
7. Coach Review: Routes to the coach for approval/editing

Safety Guardrails

Sasha has multiple layers of safety:

• Content moderation: Checks AI output for harmful, biased, or clinically inappropriate content
• Escalation triggers: Detects self-harm language, crisis indicators, or therapy-territory content and alerts the coach
• Coaching language enforcement: Ensures Sasha uses coaching language (not clinical/diagnostic language)
• Confidence thresholds: Low-confidence insights are flagged for careful review
• Rate limiting: Prevents runaway API costs and data access
• Emergency off-switch: Feature flag can disable Sasha in seconds

The "No Diagnosis" Rule

Sasha MUST NOT:

• Provide clinical diagnoses (e.g., "You have anxiety disorder")
• Prescribe treatment (e.g., "You should take medication")
• Act as a therapist (e.g., "Let's explore your childhood trauma")
• Override the coach's professional judgment

Sasha CAN:

• Observe patterns ("I notice you've mentioned feeling overwhelmed in 3 of your last 4 sessions")
• Suggest coaching questions ("The coach might explore what triggers the feeling of being overwhelmed")
• Flag concerns ("This content may be approaching therapy territory — coach review recommended")
• Provide psychoeducational context ("Research on cognitive load suggests...")

Continuous Improvement

Sasha gets better over time through:

• Coach feedback: When coaches edit or reject insights, that feedback improves future prompts
• User feedback: Coachee ratings (helpful/not helpful) are tracked
• Red-team testing: Regular adversarial testing of the AI system
• Prompt iteration: A/B testing different prompt templates with outcome tracking

Technical Spec

---

Emergency Off-Switch

If Sasha produces harmful output or a security incident is detected, the kill switch can disable all AI processing in under 5 seconds. This is a non-negotiable safety requirement. The feature flag is cached in Redis with a 5-second TTL, meaning changes propagate globally within one TTL cycle.

No-Training Guarantee

Every API call to Anthropic and OpenAI includes the no-training flag. This is contractually reinforced and technically enforced. Our data is NEVER used to train foundation models. This is verified in quarterly provider compliance reviews.

Prompt Versioning

All system prompts and prompt templates are version-controlled in the codebase (not stored in the database). Changes to prompts require pull request review, A/B testing, and safety review before deployment. Prompt versions are logged with every LLM call for reproducibility.