Coinbase
Coinbase
Coinbase
Customer Support Redesign
Customer Support Redesign
Customer Support Redesign
Overview
Overview
This project came out of my work with NYU Stern's Blockchain & Fintech Club, where our team partnered directly with Coinbase's Customer Experience group. Coinbase asked us to look into their support chatbot and figure out why users still didn't trust it, even though it handled scale well. This question led to comprehensive research and product engagement, from user interviews to a final presentation with their CX team.
This project came out of my work with NYU Stern's Blockchain & Fintech Club, where our team partnered directly with Coinbase's Customer Experience group. Coinbase asked us to look into their support chatbot and figure out why users still didn't trust it, even though it handled scale well. This question led to comprehensive research and product engagement, from user interviews to a final presentation with their CX team.
We ran user interviews to map out where the chatbot broke down for real customers, from account recovery to verification to escalation with a human agent. That research surfaced five recurring pain points: response quality, personalization, transparency, instruction actionability, and the handoff between bot and agent. I focused heavily on the response redesign, proposing a layered system with a one-sentence summary up front and an actionable list of numbers and steps that feel more direct, which gives the customer confidence in the chatbot, compared to the previous approach of listing a jumble of potential reasons or solutions for the issue raised.
We ran user interviews to map out where the chatbot broke down for real customers, from account recovery to verification to escalation with a human agent. That research surfaced five recurring pain points: response quality, personalization, transparency, instruction actionability, and the handoff between bot and agent. I focused heavily on the response redesign, proposing a layered system with a one-sentence summary up front and an actionable list of numbers and steps that feel more direct, which gives the customer confidence in the chatbot, compared to the previous approach of listing a jumble of potential reasons or solutions for the issue raised.
Tools Used
Figma
Design Tool
Firefly
AI Design Tool
Nano Banana
AI Design Tool
Claude
LLM
Tools Used
Figma
Design Tool
Firefly
AI Design Tool
Nano Banana
AI Design Tool
Claude
LLM
Tools Used
Figma
Design Tool
Firefly
AI Design Tool
Nano Banana
AI Design Tool
Claude
LLM
Created
Created
2026
Process
Analyzing Pain Points & Competitive Analysis
Each person on the team researched across chatbots from similar companies like Gemini, Kraken, and Robinhood, and then we cross-compared our personal observations to compile what worked and what didn't. That gave us a baseline for what users actually found helpful in a support chatbot versus what left them frustrated and a clearer sense of where Coinbase's chatbot could be considered more effective and lead to fewer human escalation cases.
From there, the ten of us split the interview findings into five categories: response quality, personalization, transparency, instruction actionability, and chatbot-to-agent handoff. My responsibility was more focused on analyzing response quality. I talked to real users and simulated actual anecdotes I had heard from user interviews, such as a user stuck in an authentication loop or a user who experienced fraud. Every recommendation we provided had to work on both sides, cutting friction for the user while being effective enough to prevent human escalation for Coinbase's support team.
Personalized Support Engine
This diagram is where the reimagined chatbot simulation comes together. We built the system around a synthetic user agent that mimics real Coinbase customer personas, feeding in things like account history, transaction data, and prior support context. That agent talks to a Coinbase Virtual Assistant we modeled, and every exchange gets scored by an evaluation layer combining an AI evaluator with human reviewers, so no matter what edge case gets thrown at it, we can see exactly where the bot holds up and where it breaks.
The goal was to stress-test the exact pain points we found in interviews, not just theorize about them. We ran the same failure modes we saw with real users, like the authentication loop and the account-verification stall, through the simulated agent to see if the redesigned responses actually fixed the flow. Above that, the evaluation layer tracks insights and improvements over time, so the system isn't just a one-time fix. The system is meant to keep flagging edge cases and prompt-routing failures as Coinbase's product evolves.
Personalized Support Engine
This diagram is where the reimagined chatbot simulation comes together. We built the system around a synthetic user agent that mimics real Coinbase customer personas, feeding in things like account history, transaction data, and prior support context. That agent talks to a Coinbase Virtual Assistant we modeled, and every exchange gets scored by an evaluation layer combining an AI evaluator with human reviewers, so no matter what edge case gets thrown at it, we can see exactly where the bot holds up and where it breaks.
The goal was to stress-test the exact pain points we found in interviews, not just theorize about them. We ran the same failure modes we saw with real users, like the authentication loop and the account-verification stall, through the simulated agent to see if the redesigned responses actually fixed the flow. Above that, the evaluation layer tracks insights and improvements over time, so the system isn't just a one-time fix. The system is meant to keep flagging edge cases and prompt-routing failures as Coinbase's product evolves.
Personalized Support Engine
This diagram is where the reimagined chatbot simulation comes together. We built the system around a synthetic user agent that mimics real Coinbase customer personas, feeding in things like account history, transaction data, and prior support context. That agent talks to a Coinbase Virtual Assistant we modeled, and every exchange gets scored by an evaluation layer combining an AI evaluator with human reviewers, so no matter what edge case gets thrown at it, we can see exactly where the bot holds up and where it breaks.
The goal was to stress-test the exact pain points we found in interviews, not just theorize about them. We ran the same failure modes we saw with real users, like the authentication loop and the account-verification stall, through the simulated agent to see if the redesigned responses actually fixed the flow. Above that, the evaluation layer tracks insights and improvements over time, so the system isn't just a one-time fix. The system is meant to keep flagging edge cases and prompt-routing failures as Coinbase's product evolves.
Conclusion
Working on Blitz was honestly one of the most fun projects I've done. I got to spend a lot of time on prompt engineering and figured out pretty quickly that getting a good image is way more about how specific you are than how creative the prompt sounds. I learned when to use Firefly versus Gemini, when to regenerate from scratch versus edit, and how much faster the whole process gets once you know which tool fits which job.
The one thing I'd flag is that the images aren't perfectly consistent across the project. Some details shift from shot to shot since each image is generated separately. That said, when you look at all of them together, they paint a pretty clear picture of what Blitz would actually feel like, which is the whole point of a concept like this.
Coinbase's CX team responded well to the presentation, and ideas like the layered response system and the persistent status tracker lined up closely with directions they were already considering internally. This validation confirmed the pain points from our interviews weren't isolated complaints but were, in fact, patterns Coinbase had already noticed. It also pointed to a clear next step: running the simulation against live production data instead of synthetic personas to catch edge cases that interviews alone couldn't surface. Overall, this was a great opportunity, and I enjoyed working with the Coinbase team on this project!
Link to Full Presentation
The link below is a more detailed slide deck, compiling our user research, competitive analysis, pain points, and next steps that we presented to Coinbase's CX team:
https://drive.google.com/file/d/1PYuauvqKjVzRvrpy-ksr1DlUB95jPKEe/view?usp=sharing
Conclusion
Working on Blitz was honestly one of the most fun projects I've done. I got to spend a lot of time on prompt engineering and figured out pretty quickly that getting a good image is way more about how specific you are than how creative the prompt sounds. I learned when to use Firefly versus Gemini, when to regenerate from scratch versus edit, and how much faster the whole process gets once you know which tool fits which job.
The one thing I'd flag is that the images aren't perfectly consistent across the project. Some details shift from shot to shot since each image is generated separately. That said, when you look at all of them together, they paint a pretty clear picture of what Blitz would actually feel like, which is the whole point of a concept like this.
Coinbase's CX team responded well to the presentation, and ideas like the layered response system and the persistent status tracker lined up closely with directions they were already considering internally. This validation confirmed the pain points from our interviews weren't isolated complaints but were, in fact, patterns Coinbase had already noticed. It also pointed to a clear next step: running the simulation against live production data instead of synthetic personas to catch edge cases that interviews alone couldn't surface. Overall, this was a great opportunity, and I enjoyed working with the Coinbase team on this project!
Link to Full Presentation
The link below is a more detailed slide deck, compiling our user research, competitive analysis, pain points, and next steps that we presented to Coinbase's CX team:
https://drive.google.com/file/d/1PYuauvqKjVzRvrpy-ksr1DlUB95jPKEe/view?usp=sharing
Conclusion
Working on Blitz was honestly one of the most fun projects I've done. I got to spend a lot of time on prompt engineering and figured out pretty quickly that getting a good image is way more about how specific you are than how creative the prompt sounds. I learned when to use Firefly versus Gemini, when to regenerate from scratch versus edit, and how much faster the whole process gets once you know which tool fits which job.
The one thing I'd flag is that the images aren't perfectly consistent across the project. Some details shift from shot to shot since each image is generated separately. That said, when you look at all of them together, they paint a pretty clear picture of what Blitz would actually feel like, which is the whole point of a concept like this.
Coinbase's CX team responded well to the presentation, and ideas like the layered response system and the persistent status tracker lined up closely with directions they were already considering internally. This validation confirmed the pain points from our interviews weren't isolated complaints but were, in fact, patterns Coinbase had already noticed. It also pointed to a clear next step: running the simulation against live production data instead of synthetic personas to catch edge cases that interviews alone couldn't surface. Overall, this was a great opportunity, and I enjoyed working with the Coinbase team on this project!
Link to Full Presentation
The link below is a more detailed slide deck, compiling our user research, competitive analysis, pain points, and next steps that we presented to Coinbase's CX team:
https://drive.google.com/file/d/1PYuauvqKjVzRvrpy-ksr1DlUB95jPKEe/view?usp=sharing