eplanet AI Systems — RAG Financial Assistant

The Challenge

eplanet Brokers — a live fintech trading platform with 20,000+ active users — needed to move from scattered AI experiments to reliable, production-grade AI systems. The mandate: ship AI that actually works in a regulated financial environment, serves thousands of users daily, and doesn't hallucinate.

I was brought in as the person who sits between executive vision and engineering reality: responsible for AI strategy, product execution, and the full design system modernization.

My Role

AI Product Engineer & Design Systems Lead — responsible for:

• Defining the AI roadmap and system architecture
• Full UX/UI design through to deployment
• RAG pipeline engineering and hallucination mitigation
• Autonomous agent design and workflow architecture
• Figma design system (tokenized, cross-product)

The Work

1 — RAG-Powered Financial Intelligence

The problem: Traders needed fast, reliable access to market data, platform documentation, and compliance information. Manual search was slow. Generic chatbots hallucinated.

What I built: A retrieval-augmented generation assistant that gives users secure, real-time access to verified financial information. I architected the full system — from retrieval logic and vector database design to the user interface and trust signals that make traders actually believe the outputs.

Key design decisions:

• Chunking strategy optimized for financial content (preserving numerical context)
• Source citation in every response — "show your work" as a trust mechanic
• Fallback escalation to human support for edge cases
• Interface designed for high-stress, high-stakes decision environments

Result: Live in production, serving 20,000+ active users daily.

2 — Autonomous Operations Mesh

The problem: Client onboarding, eligibility verification, and CRM data entry were consuming enormous manual hours — error-prone and slow.

What I built: An agentic pipeline that handles the full onboarding sequence autonomously: data collection, eligibility checks, CRM writes, notification routing, and exception escalation.

Architecture: n8n orchestration + Python services + CRM API integration, with human-in-the-loop checkpoints for regulatory compliance.

Result: ~70% reduction in manual administrative overhead with measurably faster client response times.

3 — AI Content & Market Analysis Pipeline

The problem: The marketing team needed daily market news, analysis, and social content at a volume impossible to produce manually. Fully automated content carried compliance and quality risks.

What I built: A Human-in-the-Loop content pipeline — AI drafts at scale, human editors approve and refine, automated publishing handles distribution. The interface was designed so editors could move fast without losing editorial control.

Result: Marketing output scaled significantly while editorial quality standards were maintained.

4 — Design System Modernization

Scope: Full tokenized Figma design system covering the trading platform, CRM, mobile app, and marketing site — as the company expands its AI offering.

Approach: Component audit → token architecture → component library → developer handoff documentation. Built to support multiple product teams working in parallel without visual drift.

Outcomes

What I Learned

The hardest part of AI implementation isn't the model. It's the interface design — specifically, designing systems that users trust. A technically correct RAG system that users don't believe is worthless. Every decision about how information is presented, sourced, and explained is an adoption decision.

The second hardest part is designing for failure. Every AI system needs explicit human escalation paths, not as a fallback, but as a designed feature.