PROJECT PROFILE — AI-Powered Fabric Sourcing Platform

At a glance: 70,000+ fabrics matched visually · < 1 s per query (vs 8–14 min) · 733 signals per fabric · 70 tunable knobs · 4 persona cockpits.

1. Snapshot

The manufacturer's sourcing teams narrowed a 70,000+ fabric catalogue using a slow spreadsheet-and-thumbnail workflow — 8–14 minutes per query, with no visual AI. Agentryx designed and is delivering an on-prem AI sourcing platform that matches fabrics visually in under a second, pairing the CLIP visual backbone with a proprietary, explainable 4-pillar engine (Colour · Pattern · Texture · Finish) and auto-extracted metadata. It is delivered in phases — a visual foundation, a capture + engine uplift, and an on-prem AI Hub of modular capsules — with all data kept inside the manufacturer's firewall.

2. The Challenge

The manufacturer held 70,000+ fabrics across 100–200 metadata fields that were mostly blank or decades old, and had no visual AI — so every "find a similar fabric" was human-eyed and slow. The live image library consisted of Photoshop-rendered exports with zero camera EXIF, which structurally caps achievable accuracy on texture and finish. The work had to satisfy hard constraints: strict on-prem / data sovereignty (global-brand confidentiality), a conservative enterprise buyer, sub-second performance at full catalogue scale, and coexistence with a legacy content system (FCMS).

3. What We Built (the solution)

Phase 1 — Foundation: sub-second visual match on the full 70K library; the proprietary 4-pillar DNA engine on a CLIP backbone; a SpotLight discovery mode; a live calibrator; per-match pillar bars that show why each fabric ranked (explainability); saved searches, Excel/CSV/PDF export, and role-based access; plus continuous auto metadata backfill that fills 50–100 fields per fabric. Phase 2 — Capture + Engine Uplift: studio capture stations and the engine enhancements they unlock — true colour calibration, cross-polarised finish separation, motif-scale measurement, and textile-tuned embeddings. Phase 4 — On-prem AI Hub: a set of modular capsules (natural-language sourcing, tech-pack parsing, RSL compliance pre-check, Customer 360, duplicate detection), each reusing the foundation. The platform presents one engine through four persona cockpits — Designer, Sourcing, Buyer, Admin.

4. Architecture & How It Works

A multi-engine, layered design.

• Experience: React / Next.js — four persona cockpits and a live configurator (70 knobs).
• API & services: FastAPI (Python), RQ background jobs, JWT / SSO / RBAC.
• Matching engines: an OpenCLIP / CLIP vision transformer produces a 512-dimension fingerprint; on top, the proprietary 4-pillar engine extracts 733 signals — Colour (ΔE2000 / LAB clusters, Pantone overlap), Pattern (Gabor / LBP / FFT periodicity), Texture (full GLCM), Finish (specular geometry) — exposed through 70 tunable knobs.
• Retrieval & data: a hybrid FAISS HNSW + pgvector index, PostgreSQL, MinIO — sub-second at 2× live scale (150K).
• On-prem infrastructure: a 4-machine cluster (application, data, GPU AI inference, DR/staging); the Phase-4 AI Hub adds a self-hosted LLM stack (~32B reasoning + 7B vision + 3B composer) served via vLLM.

Key design decisions (with rationale): on-prem for data sovereignty; CLIP as a backbone with a proprietary explainable layer on top (rather than a black-box similarity score); and deferring the GPU inference box to Phase 4 to keep the entry cost low.

5. Technology Stack

• AI / ML models: OpenCLIP / CLIP (ViT) embeddings, zero-shot & custom classifiers, contrastive fine-tuning, on-prem LLMs (~32B / 7B / 3B), RAG.
• Vision / CV: image similarity, semantic segmentation, GLCM texture, colour science (HSV, CIELAB, K-means, ΔE2000), Gabor / LBP / FFT pattern analysis, specular / finish analysis, motif-scale measurement.
• Vector & data: FAISS (HNSW), pgvector, PostgreSQL, MinIO; hybrid vector + metadata search.
• Backend: FastAPI, Python, RQ (Redis Queue).
• Frontend: React, Next.js / Vite.
• Infra & deployment: on-prem, Docker, vLLM GPU inference, 4-machine architecture, JWT / SSO / RBAC.
• Hardware: Sony α7R V capture stations (D65 cross-polarised lighting, X-Rite ColorChecker); GPU inference servers (RTX A6000 / L40S class).

6. Capabilities & Expertise Demonstrated (Capability → Evidence)

• Computer-vision engineering → built a proprietary 733-signal, 4-pillar fabric-matching engine with explainable per-pillar scoring.
• Applied colour science → ΔE2000 / CIELAB calibration pipeline anchored on in-frame ColorChecker capture.
• Vector search at scale → sub-second hybrid FAISS + pgvector retrieval over 70K SKUs, engineered to 150K.
• On-prem GenAI / LLM deployment → self-hosted multi-model stack (reasoning + vision) via vLLM, with zero data egress.
• Domain modelling & data strategy → textile-specific taxonomy and auto-extraction of 50–100 metadata fields.
• Enterprise integration → Penelope CAD ingestion, legacy FCMS, and an ERP / CRM-ready architecture.
• Solution & cost architecture → a phased four-stage roadmap, a man-month cost model, ROI modelling, and capture-hardware specification.
• Product & UX → four persona cockpits and a live, governable configurator.

7. Hard Problems Solved / Innovations

• The "domain gap" — messy real-world phone uploads versus a pristine studio catalogue — solved with semantic segmentation, configurable score boundaries and structural weighting, so the engine fails gracefully rather than returning confusing matches.
• The black-box limitation of generic visual search — addressed by building a proprietary, explainable 4-pillar engine (70 tunable knobs, visible per-pillar scoring) instead of relying on CLIP's opaque 512-vector, giving merchandisers both control and the "why".
• A rendered-not-photographed library capping accuracy — solved with a studio-capture pipeline featuring cross-polarised finish separation and a contrastive render-vs-studio fine-tuning flywheel.
• On-prem GenAI economics — selected a ~32B-class reasoning model that matches a 72B on structured tasks at roughly half the VRAM.

8. Outcomes & Impact

• Measured: replaces an 8–14-minute manual visual query with sub-second matching; Phase 1 foundation built and demonstrated as a working POC.
• Modelled (to validate in Phase 1 discovery): per-pillar accuracy gains from studio capture (overall ~46% → up to ~92%; Finish up ~8.2×) and a Phase-4 steady-state value of ₹10–20 Cr/year — against a committed, phased programme of ~₹2.77 Cr.

9. Roadmap & Future Development

A land-and-expand model — each phase is a standalone, signable scope; stop after any one and the value already delivered stands.

• Phase 1 — Foundation (Months 1–4) · built / in delivery: sub-second visual match (70K), 4-pillar DNA engine, SpotLight discovery, live calibrator, per-match pillar bars, saved searches + export, role-based access, auto metadata backfill.
• Phase 2 — Capture + Engine Uplift (Months 5–9) · next: studio capture stations, ΔE2000 colour calibration, cross-polarised finish separation, motif-scale measurement, textile-tuned embeddings, Enhanced engine + strictness dial, A/B compare, Mood Board, persona presets, full configurator.
• Phase 3 — FCMS Consolidation · parked (scoped after discovery): legacy FCMS consolidation, hybrid image + faceted search, migration tooling.
• Phase 4 — On-prem AI Hub (Months 10+, two waves) · planned: Wave 1 — Duplicate Catcher, a natural-language sourcing assistant, Tech-Pack Parser; Wave 2 — RSL compliance pre-check, Customer 360, Weekly Sourcing Report; self-hosted LLM stack served via vLLM.
• Phase 5 — ESG Monitoring & Management · future track (separate funded): compliance lifecycle registry, certificate expiry & audit alerts, buyer-readiness scoring, ESG dashboard & reporting.