Scene Editor & Illustration Agent
Prompt-driven illustration with character continuity
Each scene carries a structured image description with injected variables (character traits, settings, visual style). The illustration agent calls gemini-2.5-flash-image via OpenRouter with a multimodal character reference sheet, ensuring visual consistency across all 20 scenes.
Semaphore-bounded asyncio.gather runs up to 5 scene generations in parallel. If OpenRouter rejects a reference-image request, a single retry fires with text-only context — partial failures never abort the batch.
Multi-Domain Book Generation
Same pipeline, radically different outputs
The system is fully domain-agnostic. The same agentic pipeline that generates a children's emotional literacy book produces a Pixar-style historical non-fiction illustrated narrative — no code changes, only a different prompt file and character sheet.
The framework adapts to domain through configuration, not re-engineering — a core property of well-designed agentic systems.
Multi-Model LLM-as-Judge
Automated accuracy verification across 4 models
A parallel asyncio jury grades every scene: accuracy score (0–100), editorial dimensions (age appropriateness, political risk, reading level, mood, style), and inter-model agreement.
Confidence score blends mean grade with inter-model spread to detect specification gaming. Adversarial probe injection tests robustness.
Partial model failures surface as null grades — the jury never aborts. Models: GPT-4o-mini, Gemini 2.0 Flash, Claude 3.5 Haiku, Mistral Large.
Library & Storyline Orchestration
Multi-book, multi-user, concurrent
Each book is an independent storyline with its own scene graph, prompt file, character reference, and assembly block sequence. Multiple books can be in active generation simultaneously, each using an isolated SessionLocal() to prevent ORM conflicts across concurrent commits.
Per-book actions — Editor, Assemble, Accuracy, Promo.
Auto-Generated Promotional Page
From assembled book to market-ready shareable page — no design tool, no manual layout.
AI Cost Attribution by Call Type
Full observability into per-operation model spend
Every AI call is logged to an ai_calls ledger: model, call_kind, cost_usd, latency, user, and storyline linkage. This granularity enables ongoing pricing model validation, per-user profitability analysis, and model benchmarking across call types.
Knowing the cost and performance of each agent operation — not just the total — is what allows a system to be optimised, not just monitored.
Operational & Financial Tracking
Full visibility into revenue, AI spend, credit activity, and live data queries — built in.
