E*Trade Chatbot

The Problem: High Call Volumes, Frustrated Users

E*Trade's contact center was flooded with routine questions (e.g., "How do I reset my password?"). Users waited on hold for simple tasks, and the company needed to cut costs without sacrificing service. Our challenge: create a self-service tool that felt helpful, not robotic.

What We Built

A virtual assistant (powered by Kore.ai) that handles 300+ common financial queries, from account updates to trade confirmations. Key features:

• 24/7 live chat: Instant answers without waiting.
• ADA compliance: Screen-reader-friendly design and keyboard navigation.
• Smarter conversations: Improved natural language processing (NLP) to understand slang or typos (e.g., "How 2 change my email?").
• Knowledge graph updates: Added answers to recurring questions users asked but weren't covered.

Our accessibility features were a core focus, ensuring that every user could navigate the tool regardless of ability. We implemented high-contrast options, keyboard shortcuts, and ensured all elements were properly labeled for screen readers.

How It Works

• User asks a question: The chatbot checks the knowledge graph for matches.
• NLP interprets intent: Even if phrasing is unusual, the system identifies the core request.
• Fallback to human agents: Complex issues get routed to live support seamlessly.

Behind the Scenes

Building this required:

• Collaboration with annotators: To label thousands of user queries for training the NLP model.
• ADA audits: Working with accessibility experts to fix contrast ratios, alt-text, and keyboard controls.
• Gap analysis: Reviewing chat logs to find questions the bot couldn't answer, then updating the content (e.g., adding FAQs about tax forms).

The biggest hurdle? Making financial jargon (like "cost basis" or "settlement date") sound simple. We rewrote 70+ answers with plain language and examples.

Behind the Scenes: UX, Copy, and Collaboration

1. Crafting the Chat Personality

• Tone: Avoided jargon (e.g., "settlement date" became "the day your trade finalizes").
• Error handling: Replaced "Invalid input" with "Let me try that again—could you rephrase?"
• Testing: Ran A/B tests on response styles (formal vs. casual) — users preferred short, upbeat replies.

We found that humanizing the bot's responses significantly improved user satisfaction. By creating a conversational tone that recognized frustration and acknowledged mistakes, users were more likely to continue using the chatbot instead of calling customer service.

2. Designing the Interface

• Wireframes: Mocked chat bubbles, button placements, and loading states in Figma.
• Accessibility: Worked with visually impaired users to test screen-reader flow.
• Micro-interactions: Added subtle animations (e.g., typing indicators) to reduce frustration.

Interface iterations focused on visual clarity and reducing cognitive load. Each UI element was designed to guide users toward successful interactions and provide clear paths for complex requests.

3. Building the Frontend

• Kore.ai integration: Built reusable modules for common tasks (password resets, balance checks).
• Knowledge graph gaps: Added 200+ missing FAQs by analyzing chat logs (e.g., "How do I update my W-9?").
• Fallback logic: Designed "Sorry, I'm still learning!" messages with quick exit buttons to human agents.

When the chatbot couldn't resolve an issue, we ensured a smooth handoff to live agents. Context was preserved, eliminating the need for customers to repeat information.

What We Learned

• Users prefer "quick fixes": 80% of chats were resolved without transferring to humans.
• Language evolves: The bot needed monthly updates to keep up with slang.
• Accessibility isn't optional: Testing with screen-reader users revealed flaws we'd missed.

Results

• 13.4 Million hits per month
• Average chat volumes of 33.6k per month
• Conversations per page:
  • Service Center - 20k+
  • Move Money - 13k+
• Faster resolutions: Average chat time: 1.3 minutes (vs. 8+ minutes on calls).
• Inclusive design: Achieved ADA compliance, making support accessible to users with disabilities.
• Reduced calls: 20% drop in contact center volume within the first month.
• Higher satisfaction: 88% of users rated the chatbot "easy to use."
• Faster resolutions: 90% of password resets happened in under 2 minutes via chat.

Project Role

Software Engineer (Backend Integration, UI Requirements, Accessibility Lead)

This project taught me that good tech doesn't just solve business problems—it removes everyday frustrations. Sometimes, the most impactful work is invisible: like a chatbot that lets people skip the hold music.