Skip to main content

​
Roadmap

Brain is being built one workflow at a time, deployed to real users on real data, then iterated. Every shipped workflow includes tracing, correction logging, and a changelog entry — that bar applies from day one and never gets dropped. This page is the commitment view. Status indicators tell you what’s live, what’s in development, what’s planned next, and what’s further out.
Status legend:
  • 🟢 Live — running with real users on real data
  • 🟡 In Development — being built now
  • ⚪ Planned — committed, next up
  • 🔵 Future — backlog, after we ship the foundations

​
Phase 1 — Anchor Feature: Tech Pack → BOM

🟡 In Development The first capability is the highest-pain manual work in the studio. Mai spends hours on every new tech pack: silhouette, fabric panels, buttons, zippers, trims, interlining, thread, packaging, labels, hidden construction materials. Brain reads the tech pack and produces a first-draft BOM with confidence scores. Why this first: highest pain, highest ROI, hardest to fake with a thin wrapper. If this works, the rest of the framework gets built around a feature people already use. If it doesn’t, we discover that early and adjust. Built by: Ku Kue (dedicated agent track) Includes:
  • PDF parsing of tech packs
  • Excel parsing of POM and measurement sheets
  • Vision analysis of garment images
  • Category classification
  • Visible material extraction
  • Hidden material template insertion (jacket, pants, dress templates)
  • BOM draft written to database with AI First Draft status
  • Confidence scoring per BOM line
  • Correction logging for every human edit
Ships with: tracing, correction log table, first changelog entry. See Tech Pack → BOM for the full spec.

​
Phase 2 — Framework, Email Intake, Style Creation

⚪ Planned (in parallel with Phase 1) While Ku Kue builds the BOM agent, the framework gets built around it: scoped chat UI, role and permission system, Inbox approval cards, task system, email ingestion, style creation workflow, duplicate detection. Includes:
  • Scoped assistant wrapper on deer-flow runtime
  • Role and permission gating
  • Inbox surface for approvals and AI messages
  • Task system (human tasks + AI monitoring tasks)
  • Email webhook ingestion
  • Email tagging and classification
  • Attachment detection
  • Style creation workflow (manual + automated)
  • Email-to-style flow
  • Client creation workflow
  • Duplicate style check
  • Draft preview and approval cards
  • Promotion of approved drafts into CutMake
Built by: Thai (framework track)

​
Phase 3 — True Cost and Quotation

⚪ Planned Once BOM drafts are reliable enough to review, the next workflows are true cost calculation and quotation generation at fixed markup. Includes:

​
Phase 4 — Oversight and Manager Intelligence

⚪ Planned Once the intake-to-quotation path is shipping, the next layer is visibility. Managers need to know which RFQs are pending, delayed, quoted, cold, or at risk. They need to ask “what’s happening with this style?” and get a complete, source-linked answer. Includes:

​
Phase 5 — Pricing Intelligence

🔵 Future Once we have enough history (sent quotations, accepted/rejected outcomes, negotiation threads), the next layer is helping the team price better. Includes:
  • Comparable style search (external + internal)
  • Retail-to-FOB estimation
  • Historical quotation analysis
  • Client margin profiles (Jaspal vs. Lulus vs. ASOS)
  • Negotiation history summarization
  • Suggested quote ranges with rationale
This is the layer where data accumulated over phases 1-4 pays off. We don’t build it early because the eval datasets and historical signals don’t exist yet.

​
What ships parallel, always

These are not phases. They ship alongside every feature.
  • Tracing through the chosen observability tool. See Observability.
  • Correction logging in the Brain database. Every human edit of an AI draft is recorded for measurement and improvement.
  • Changelog entry in this documentation. Every shipped feature is documented in the Changelog.
  • Auditability — every output traces to source artifacts.

​
What we won’t do

This is the equally important roadmap. Things we’ve explicitly decided against.
  • No general-purpose AI chat. Scoped workflows only.
  • No autonomous emails. Every send requires human approval.
  • No autonomous record creation in CutMake. Drafts only; humans promote.
  • No agent-only design partners. Mai and Zean drive the product; we don’t scale to more users until those two are satisfied.
  • No mock data validation. Every workflow is tested against real RFQs.
  • No skipping the changelog. If it ships, it’s documented.