SalesOS is built as multi-tenant AI infrastructure for sales skill intelligence — not a chat toy, but a structured feedback and measurement system.
Claude claude-sonnet-4-6 acting as enterprise buyer personas. System prompts enforce role, difficulty, objection style. Context window maintains session memory.
Separate Claude instance analyzes rep transcript. Outputs structured JSON against 5 skill dimensions. Validated with Zod schema before storage.
System prompts prevent buyer from being helpful, maintain character, enforce response length, simulate real objection patterns.
Row Level Security (RLS) enforced at the database layer via Supabase PostgreSQL policies. Each organization has hard data boundaries — reps can only see their own sessions, managers can only see their org's team data.
Scoring responses are validated against a strict Zod schema before being accepted. Malformed or incomplete JSON is rejected — ensuring data integrity across all skill dimensions.