Skip to content

AI design scout

DesignBoard

Scouts the live web for emerging lab-grown jewelry designs, scores each one on evidence, and turns the best into manufacturing blueprints.

Role
Sole builder. Discovery, scoring, agent pipeline, validation, and interface.
Status
In development
Year
2026
How it works, end to end. The screens inside are real captures; every command and stage is the project's own.

The problem

A jewelry team needs to know which designs are rising before everyone makes them, and a factory needs specifics, not mood images. Trend boards give neither: no evidence behind the pick, and no measurements to build from.

What I did

Built the live scout, the evidence scoring, and a four-phase agent job that writes a factory brief, plus the code that audits it.

How it works

Discovery reads retailer and marketplace pages and keeps only diamond-led, non-basic designs with early momentum and low observed saturation, filtered for everyday wear, light gold and practical CVD manufacturing. Each card can then run a resumable agent job (observing, drafting, reviewing, validating) whose output must fit a 13 KB JSON schema and pass a deterministic audit that refuses measurements the photo cannot support.

A DesignBoard card showing design grammar, CVD manufacturing direction and evidence scores.
Each card: the design grammar, an original CVD direction, and the evidence behind the score. Retailer photos blurred.
Three scouted design cards with category, source and watchlist score.
Scored candidates, filtered by category.

The hard parts

  1. 01

    Keeping a model honest about measurements: anything the photo cannot show is downgraded to a stated assumption, never printed as a fact.

  2. 02

    Long agent jobs that survive restarts: atomic state files, resumable phases, and tracked child processes that are stopped cleanly when interrupted.

  3. 03

    Separating inspiration from copying: every card must state an original lab-grown adaptation, and natural-diamond sources stay labelled as reference only.

What changed

  • 38 emerging candidates on the current board, each with evidence links and a score.
  • 11 finished manufacturing blueprints from real runs, each printable as a factory brief.

Built with

  • React
  • TypeScript
  • Vite
  • Ajv
  • Node
  • Codex CLI