Zizta

The problem

AI content tools produce confident, generic drafts. Teams could not tell which output was actually good before publishing it.

Business challenge

Make quality measurable: every generated brief needed a score a human could trust and argue with.

Research

Built a labelled set of 600 briefs graded by editors, then tuned an evaluation rubric until the model agreed with editors 90% of the time.

Planning

A multi-stage pipeline - retrieve, draft, critique, revise, score - with each stage cached, observable and independently replaceable.

Architecture

• Queue-driven generation pipeline with per-stage retries
• Vector retrieval over crawled SERP and client corpus data
• Model routing: cheap models for extraction, strong models for critique
• Full prompt/response logging for evaluation and cost attribution

Database design

PostgreSQL with pgvector for embeddings, plus a run table capturing every stage's input, output, tokens and cost.

API flow

1. 1 User submits a keyword and target audience
2. 2 Retrieval stage assembles SERP context and internal corpus chunks
3. 3 Draft → critique → revise loop runs until the score plateaus
4. 4 Final brief and its score card are returned and stored

Implementation

• Started with evaluation harness before writing a single prompt
• Streaming UI so users watch reasoning instead of waiting
• Per-workspace cost caps to keep unit economics predictable

Key features

• SERP-grounded content briefs
• Automated critique with a transparent score card
• Workspace-level cost and usage controls
• Streaming generation UI

Performance

• Model routing cut token spend without changing output scores
• Cached retrieval reduced repeat-run latency by 3x

Lessons learned

• Build the evaluation before the feature. It changes every later decision.
• Cheap models do 70% of the work if you route deliberately.