Abbeal
bnp logo

Tier-1 bank · Paris

BNP Paribas: Reference Book PO, from React/Redux to product AI agents.

Three Abbeal engineers at the core of the PO Marketplace. React/Redux/Node platform initially, now augmented with a product RAG, Claude agents for PM assistance, and an event-driven Kafka layer to scale.

KPI

RAG

PO product catalog

Duration

Engagement multi-année

Team

3 engineers

Hub(s)

Paris

Next.js 16Claude SonnetLangGraphpgvectorAWS BedrockApache Kafka

BNP Paribas. Reference Book PO Marketplace. Three Abbeal engineers embedded on the product side to build the platform that structures the product catalog and PO workflows of a tier-1 bank.

Starting point (2018-2019)

Sebastien, Raphael and Ulric embedded in the Reference Book team. Initial stack React + Redux + Node, on-prem deployment, integration with core banking. Mission: pull the PO Marketplace out of Excel hell and run it on a stable, governed, auditable web platform.

What was delivered

  • Unified PO front-end (catalog + active POs + review workflows)
  • Node API for PO step orchestration and notifications
  • Regulatory reporting and audit trail for internal controls
  • Documentation and training for internal PMs on the platform
  • Integration with existing BNP tools (SSO, AD, internal ticketing)

The stack we ship today

On the same problems today (financial product catalog, PO workflows, multi-criteria search across tens of thousands of items), here's what we assemble:

  • Next.js 16 + React 19 on the front, Server Components by default, Edge runtime on listings
  • Apache Kafka to decouple PO workflows and enable event replay
  • Product RAG on pgvector: semantic search across the catalog (10k+ products, 50+ typed attributes)
  • Claude Sonnet agents via LangGraph to assist PMs: product sheet drafting, regulatory matching, contract term wording
  • AWS Bedrock to host LLMs on the compliance side + LangSmith observability for traces
  • Auth0 + fine-grained RBAC for data scoping per desk and jurisdiction

Why it's hard (and why few succeed)

  • Fragmented product data: 6+ source systems, different conventions per desk
  • Required latency: a PM won't read an answer at > 2 seconds
  • ACPR + EBA + internal audit compliance on every agent suggestion
  • Citation required: no answer accepted without pointing to the source in the reference
  • Coexistence of on-prem Mistral (sensitive data) + Claude API (non-sensitive workflows)

What this engagement taught Abbeal

BNP was our first deep dive into financial product PO. We learned to ship in a commit-audited environment, to code for PMs who don't forgive latency, to document for 6-month internal control reviews. That's the experience we replay today on product RAG for other European tier-1 banks.

A similar case at your place?

Talk to an architect