Case Study
Building a SaaS business, with an AI product team
What started as an experiment in how far I could leverage all the new AI technology available, became a passion project, and ultimately a fully functioning product with a company built around it.
After experiencing repeated delays or total silence from local businesses when requesting quotes by email, I recognized a critical service gap: small businesses were losing money, and trust, due to poor customer response times.
This inspired me to build a customer support automation app that uses generative AI to answer incoming emails within seconds - accurately, consistently, and with contextual awareness of each business’s offerings.
The Problem
“I just wanted a quote — and they never got back to me.”
Small businesses are often overwhelmed. They don’t have the bandwidth to monitor their inboxes around the clock, but every missed message is a missed opportunity.
Pain Points:
Slow or no response to inquiries
Negative customer perceptions
Lost revenue due to lack of follow-up
Inability to prioritize urgent or emotionally charged messages
The Process
Research & Discovery
I began by using Perplexity Pro deep research mode to conduct generative user research and market analysis. This approach helped me quickly surface patterns, validate assumptions, and test early product concepts.
Research Methods:
Surveys with targeted audience segments - Small business groups via LinkIn and Reddit
Virtual expert interviews - Zoom interviews with SMB owners
Perplexity Deep Research Pro - Public dataset and industry report analysis
Defining the Solution
The final product vision became clear:
An AI-powered email response system, trained on unique product or service data, responding to customers within 60 seconds.
Key Features:
Monitors customer support inbox every 60 seconds
Uses a RAG-trained GPT to draft responses with 100% accurate product knowledge
Calculates a custom “confidence score” for each message
High-confidence → auto-reply
Low-confidence or flagged → routed to human review
Sentiment analysis detects and escalates conflict-prone messages to human managers
Concept Sketching and Wire framing
With net-new concepts I still prefer to start with pen and paper. - Once I have a rough idea, I move to Balsamiq, or in this case straight to Figma for some quick wire frames.
UI Design and Prototyping
The UI was mocked up in Figma combining the wire frame layout with the color pallet used on the website. - I then used the Figma to AI Code plugin to export the mock as static HTML and CSS.
I cleaned up Figma generated HTML and uploaded it, along with the Figma mock, into my Claude project and prompted Claude to refactor the HTML to use Tailwind for styling and build in the websockets to connect to the backend .
Tooling Decisions
My original intent was to prototype the app using no-code platforms like Bubble, Bolt, and Lovable. While these tools accelerated early exploration, they quickly hit architectural and scalability limits.
Instead, I pivoted to a hybrid approach:
Design the system, user flows and UI in Figma
Collaborate with custom Claude Sonnet to structure the Python and JavaScript
Manually code the app using traditional dev tools, Sublime text, with Claude's assistance.
Employ various LLMs leveraged as “AI dev collaborators”
This application is mostly backend processing and requires a single UI, the email dashboard, so I used Figma to AI Code to export initial HTML and then refine final code manually and with Claude Opus 4 .
LLMs as my Product Team
This was not a solo build - I collaborated with a virtual team of LLMs, each customized for, and assigned a distinct role:
Sr. Python Engineer - Claude Sonnet 4
Trained with project context and prompt engineering
Used for most Python, Flask, JavaScript, Socket.IO, and backend logic
Updated post-cutoff knowledge via RAG-fed docs
General web developer - Gemini 2.5 Flash
Used for resolving unfamiliar concepts and general architecture questions
Quick reference for Git and Terminal command line questions
Integrations Engineer and Python Consultant - Claude Opus 4
Served as my heavy hitter for tough debugging issues
knowledge source for API integrations
Used when other LLMs get stuck in a trouble shooting loop
This AI triad allowed me to:
Maintain momentum through roadblocks
Get rapid answers without diluting the core thread
Parallelize tasks and brainstorm implementation ideas
The Results
The end product is a lean, scalable, automated customer support tool that solves real-world inefficiencies. It is currently running in a headless beta on two real world environments.
Tangible Benefits:
60-second response time to most inquiries
Reduced human workload while maintaining oversight
Enhanced customer satisfaction through speed and relevance
Easy integration into existing SMB workflows