AI Agents · CATEGORY-INTELLIGENCE · Building

Category Intelligence Agent.

Category × season × geography → 40-50pp evidence-tagged trend report. Built for buying teams, not for press.

Building PLAN L4

Pipeline

7 steps

Brief → Research → Images → Build → Polish → Verify → Audit → Ship

Output depth

40-50 pp

evidence-tagged · A4 PDF

Quality checks

18 automated

post-build · regression-gated

01 · The job to be done

The pain

Trend research today = ad-hoc Pinterest scrolls + WGSN extracts + a designer's gut. Inputs are unstitched to brand DNA, evidence rarely survives the handoff to buying, and trend calls land as opinions, not signals you can audit.

What the agent does

Takes a category × season × geography brief and produces a 40-50 page PDF report: evidence-tagged trend analysis across silhouette, fabric, colour, print and detail; 5 capsule design concepts with evidence panels and AI-generated line sheets; competitive benchmarking with retailer evidence grid; risk classification (safe bet / calculated risk / test & react); and a print design brief for textile teams.

What humans still do

Buying lead approves the brief and the four review checkpoints. Designer takes the capsules into pattern-making. Agent does the research, structuring, image generation, and writing — humans decide what to act on.

02 · How it works · the closed loop

Inputs (blue) → 7-step pipeline with 4 human review checkpoints (yellow) → 40-50pp evidence-tagged report.

Brief ingestion

Buying lead supplies category · season · geography. Plus brand DNA, price tier, target consumer, occasion calendar (festive / wedding / monsoon).

Research · 3 tiers

Tier 1 runway + forecast (Vogue Runway, Tagwalk, WGSN, EDITED, Heuritech). Tier 2 retail (Myntra, AJIO, Nykaa, Net-a-Porter, Reliance Trends). Tier 3 consumer (Pinterest, Google Trends, TikTok). Every signal carries an evidence tag — RUNWAY, RETAIL, SEARCH, SOCIAL, TRADE, FORECAST.

Image generation

Line sheets, capsule illustrations, fabric textures, print swatches, cover images. GPT Image 2 + Vertex Imagen, with Fynd logo as reference for brand-mark fidelity.

Build PDF

react-pdf renders A4 portrait pages (landscape for line sheets). Structured data (palettes, capsules, ratings) + editorial prose, page-by-page designed — not HTML-to-PDF conversion.

Polish

Layout, typography, visual hierarchy. Market-specific framing — Indian climate, occasion mapping, fusion sensibility, price tier nuance.

Verify · 18 auto-checks

Image aspect ratios, banned strings, section ordering, evidence legend presence, Fynd Create branding, file size, swatch visibility, radar-chart label positioning. Regression-gated — fail one, the build doesn't ship.

Audit + Ship

Independent fashion-report-reviewer agent scores against a 5-dimension rubric (evidence rigour, content quality, visual completeness, structural compliance, editorial quality). Buying lead signs off; report ships to design team.

03 · Underlying data

Data source	Classification	Key entities
Vogue Runway / Tagwalk	Machine-Readable	Runway looks · designer attribution · season · category
Retail crawlers	Machine-Created	SKU listings · price · attributes · brand · stock signals
Pinterest / Google Trends / TikTok	Machine-Created	Search velocity · pin counts · hashtag engagement
WGSN · EDITED · Heuritech	Mixed	Forecast reports · trend velocity · colour confidence
Brand DNA brief	Human-Created	Target consumer · price tier · occasion · construction guardrails

Refresh cadence: Per-report run · research is fresh-pulled at brief time. Reports versioned with date stamp.

KPIs moved:

Time-to-trend-call (weeks → days)Evidence coverage (% claims with source tag)Capsule selection rate by buying teamReduction in trend miss-fire (in-season markdown rate on agent-recommended SKUs)

04 · Design patterns used

Wide Research

3-tier research generates many candidate trend signals before scoring narrows to a shortlist

Tool Calling

Crawlers, image generation APIs, PDF rendering, post-build check runners — all tool-mediated

Deep / Hierarchical

Slash-command orchestrator (/trend-report) routes to specialist sub-pipelines per report stage

Context Engineering

Brief + brand DNA + market frame compressed into per-section context; report-config layer enforces consistency

Evals

18 post-build auto-checks + independent fashion-report-reviewer 5-dim audit; both gate ship

Observability

Every trend call traces back to its source signal and evidence tier

05 · Evals · how we know it works

Two-layer eval. Layer 1 — 18 post-build automated checks gate the build (image ratios, banned strings, section ordering, evidence legend, file size, etc.). Layer 2 — an independent fashion-report reviewer agent scores each report against a 5-dimension rubric (evidence rigour, content quality, visual completeness, structural compliance, editorial quality), 0-100 each.

Gates and thresholds.

All 18 post-build checks pass · zero regressions
Reviewer score ≥ 75 / 100 across each of 5 dimensions
Every trend call carries an evidence tag