Neural Showrooms - Architecture

20+

OEM Integrations

15k+

Products Processed

Ingest Methods

Multi

Tenant

B2B/B2C

Modes

Data Flow (Cross-Repository)

OEM Sources

→

Neural Core Ingest

→

S3 (processed/)

→

Showrooms Scan

→

RDS + Storefront

System Components (3 Repositories)

neural-showrooms

Consumer Storefront + Catalog — Next.js 15.0.3 on Webflow Cloud (Cloudflare edge). Scans processed JSON from S3. Prisma Accelerate + Upstash Redis for edge. Downstream consumer only.

Next.js 15 Prisma 6.18 pgvector Upstash Redis Stripe

neural-showrooms-app

Admin Backend — React 18 + Lambda monorepo. Showroom lifecycle (ONBOARDING → ACTIVE → CLOSED), Attio CRM webhooks, Cognito auth, user management. Dual source-of-truth: Attio (PII), RDS (technical).

Lambda API Gateway Cognito Prisma Turborepo

neural-partners-agents

Data Ingestion Hub (Neural Core) — All upstream: FTP/SFTP/API extraction, PDF processing, web scraping, npAI enrichment. Outputs to S3 processed/.

Step Functions Lambda S3 Claude

Core Capabilities

AI Product Enrichment

npAI

Claude-powered extraction from PDFs, web scraping, and structured data. Fills gaps in dimensions, electrical specs, safety certs. Schema v2 with 200+ fields.

Claude Bedrock Step Functions

Multi-Channel Ingest

4 methods

API Push (REST), SFTP server ($7/mo Lightsail), S3 cross-account, email CSV. Per-OEM isolation, automatic S3 sync, webhook processing.

REST API SFTP S3 Email

FTP Extraction

4 data types

Pull from OEM FTP servers: products (base catalog), images (matched to SKUs), pricing (NET wholesale), inventory (stock levels). Parent inheritance supported.

Products Images Pricing Inventory

AI-Powered Search

Bedrock

Titan Embeddings V2 (1024-dim vectors) stored in pgvector. Claude Sonnet for query understanding (natural language → filters). Semantic product search with Upstash Redis caching at edge.

Titan Embeddings Claude Sonnet pgvector Upstash

Multi-Tenant Platform

showroomId

Full tenant isolation via showroomId FK. Per-showroom branding, integrations, product catalogs. Attio CRM as PII source of truth, RDS for technical config.

Tenant Isolation Attio CRM Webhooks

Lifecycle Management

5 states

ONBOARDING → ACTIVE → SUSPENDED → PENDING_CANCEL → CLOSED. Attio checkbox triggers provisioning. Automated offboarding with data export grace period.

State Machine Audit Logs Webhooks

AWS

RDS PostgreSQL Lambda S3 (3 buckets) Step Functions Bedrock API Gateway Cognito DynamoDB

Stack

Next.js 15.0.3 TypeScript 5.9 (strict) Prisma 6.18 + Accelerate Upstash Redis Tailwind CSS 3.4

AI/ML

Titan Embeddings V2 (1024-dim) Claude Sonnet (query understanding) pgvector (semantic search)

External

Webflow Cloud (Cloudflare edge) Stripe Attio CRM Brandfetch CDN Lightsail SFTP ($7/mo)

Ingest

raw → staging → processed LAK History (90-day TTL) OEM State Tracking Additive Merge

Market Readiness & Support Opportunities

Storefront Experience

65%

Functional, needs polish

Next.js 15 storefront deployed on Webflow Cloud. AI search, B2B/B2C checkout, multi-tenant routing all operational. Needs UX refinement, mobile optimization, and performance tuning before broad rollout.

Support: Frontend performance audit, UX/UI review, Lighthouse optimization

Catalog Infrastructure

70%

Architecture proven, scale untested

RDS PostgreSQL + Prisma ORM with Accelerate (Redis/Upstash) for connection pooling. Tested to 10k products. Schema supports 200+ fields per product. Multi-tenant isolation via showroomId FK.

Support: Load testing at 100k+ products, RDS scaling strategy, query optimization

npAI Enrichment Pipeline
30%
Proof of concept, not production-ready

                        Claude API-powered extraction from PDFs and web scraping. Tested on 1 OEM (~200 products). Step
                        Functions orchestration works but untested at scale. Currently on Claude API direct - needs
                        Bedrock migration for production cost/reliability.
                    

                        Priority Support: Bedrock migration, batch processing architecture, cost
                        modeling, error handling, scale testing across 100+ OEMs
                    

Automation Pipeline
10%
Manual process, no automation

                        Currently all ingest is human-triggered. No scheduled pulls from OEM FTP servers, no change
                        detection, no automatic catalog refresh. Each OEM update requires manual intervention to
                        extract, enrich, and push to showrooms.
                    

                        Priority Support: EventBridge scheduling, change detection (S3/FTP), automatic
                        enrichment triggers, catalog sync orchestration, monitoring & alerting
                    

Admin Backend (App)

15%

Scaffolded, needs significant work

neural-showrooms-app handles showroom lifecycle, user management, and integrations config. Backend Lambda/API and frontend UI both need TLC. Attio webhook sync works. May not block GA but limits self-service onboarding.

Support: API hardening, admin UI buildout, user management flows, integration configuration screens

S3 Buckets

neural-partners-ingest-raw neural-partners-ingest-staging neural-partners-ingest-processed

APIs

api.neuralpartners.ai sftp.neuralpartners.ai

DynamoDB

neural-partners-oems neural-partners-oem-state neural-partners-lak-history

Repos

neural-showrooms (storefront) neural-showrooms-app (admin) neural-partners-agents (ingest)