Claude
|
503bf2e5d7
|
feat: AI Cost, Quality & Proof OS — complete
AI Layer:
- llm_router.py: routes cheap/mid/high models, enforces daily budget, caches
- token_counter.py: estimates tokens, truncates to budget
- response_cache.py: in-memory cache with TTL per agent
- prompt_registry.py: versioned prompts with stable prefix for caching
- ai_budget.yaml: model costs, agent budgets, daily limits (10 SAR/day)
Guardrails:
- output_validator.py: blocks fake claims + prohibited actions
- cost_guard.py: prevents runaway spending
Observability:
- trace.py: trace_id, cost, latency, steps per pipeline run
Tests: ALL PASS
- 30/30 evals (100%) — 9 sectors, 30 companies
- 10/10 prohibited actions blocked
- 4/4 allowed actions verified
- 3/3 forbidden claims blocked
- 3/3 message quality checks passed
https://claude.ai/code/session_01W1rJthWDkasijTdXCfxVHs
|
2026-04-26 17:42:47 +00:00 |
|
Claude
|
18a0d95e3e
|
feat: Full Company OS — 9 new agents + scoring engine + compliance engine + evals
New agents: partnership_strategist, negotiation (10 objections), crm_revenue (16 statuses),
learning, web_search, enrichment, campaign_orchestrator, competitor_intelligence, content_strategy
New engines:
- scoring/scoring_engine.py: unified scoring with 9 sector defaults
- compliance/compliance_engine.py: channel policy + daily limits + stop words
Evals: 10/10 PASS (100%)
- Agency → email + agency_partner ✅
- Real estate → email + direct_customer ✅
- Clinic → whatsapp_warm ✅
- Ecommerce → email ✅
- Website agency → linkedin_manual + implementation_partner ✅
- Consulting → linkedin_manual + referral_partner ✅
- All: compliance=allowed, opt-out present, no prohibited actions
https://claude.ai/code/session_01W1rJthWDkasijTdXCfxVHs
|
2026-04-26 17:20:36 +00:00 |
|