A writing partner that never loses the plot

Problem

Freya is an AI writing partner for novelists. A novel is long, and the facts that hold it together are spread across hundreds of pages. A writing partner that loses track of any of them becomes one more thing the author has to check.

That makes consistency the core requirement, not a feature: every suggestion has to agree with the book so far.

Approach

As Technical Lead, I co-architected Freya’s AI layer and wrote its architecture model, the “hourglass”: many ways in, one source of truth, prose on the way out.

1. 01Channels4 ways in
2. 02Inbound3 subagents
3. 03Factsfacts only
4. 04Outbound3 subagents

1. 01Authors feed the story in through four channels: voice-first creation, guided questions, character interviews and manuscript import.
2. 02Inbound subagents extract, summarise and structure what comes in.
3. 03The result goes into a single knowledge base that holds only facts: the narrow waist.
4. 04Prose generation draws on that base, and outbound subagents generate, validate and enrich the result, which closes the loop.

Step 1 of 4

Authors feed the story in through four channels: voice-first creation, guided questions, character interviews and manuscript import.

Authors feed the story in through several channels: voice-first creation, guided questions, character interviews and manuscript import. Inbound subagents extract, summarise and structure what comes in, and the result goes into a single knowledge base that holds only facts. Prose generation draws on that base. Outbound subagents generate, validate and enrich the result, which closes the loop.

The team built the platform on that model:

• seven specialised chat agents with retrieval over the story bible;
• a guard pipeline after every generation;
• a knowledge graph with 33 relation types, including epistemic “who knows what” tracking;
• real-time co-editing on Yjs CRDTs over WebSocket.

Underneath is a FastAPI backend with more than 900 endpoints, 180+ data models and pgvector, integrating Claude (Sonnet and Opus) and OpenAI models.

My own code made that layer hold up in production:

• one model router across the agents, including content-policy-aware routing;
• per-project isolation and authorisation on chat and manuscript lookups;
• sanitised streaming errors;
• real token and cost tracking on streaming;
• blocking AI calls moved off the event loop.

Result

Many ways in, one store of facts, prose on the way out: the shape Freya’s AI layer was built on, made ready for production.

I would make the same choice again for any assistant that has to stay consistent over a long document: a narrow waist. Whatever the input, it becomes facts first, and generation reads only from those.