AI Product Design. Building a networking coach for the part of the job search everyone said was hardest
CareerCoachingPro fields FAQ and booking questions constantly. I'm one of three on the AI team building the intake bot that answers them, shares pricing, and routes people to the right service. Still in progress: fixing a hallucination issue found in testing.
Role
AI Product Designer
Client
CareerCoachingPro
Team
Karina Morera, Jim Capone,
Tariq Cilione, Founder Carmelina Pietra
Timeline
April 2026 – Present
Tools
Claude, Slack, Mural, Action Potential
Overview
The Problem
Business pricing isn't public anywhere on CCP's site, so the bot needed a deliberate answer, not a guess
Testing surfaced a real hallucination on a basic FAQ question
Three audiences, three different next steps, none of them should feel like an afterthought
The knowledge base blew past the platform's character limit, over 33,000 characters against a 12,000 cap, forcing real cuts on what the bot could hold onto
My Process
1
Built the knowledge base
directly from CCP's FAQ page and site content, nothing invented.
2
Draft Vendor Response
Onboarding, business type, services, pricing, persona, and flagged the pricing gap above for the team to decide on deliberately.
2
Designed both intake prompts
Chat and voice, personality, tone, and human handoff.
4
Hallucination Issue
Using ActionPotential testing agent and making edits to outputs to get ready for user testing.
Design Principles
Real-time transparency
Show availability immediately
Progressive disclosure
Display only relevant fields for each user type
Design layout for mobile
Make booking simple on small screens
Human guidance
Clear messaging and access to real staff
Reflections (So Far)
Where Things Stand
Knowledge base and both intake prompts are built. Hallucination issue found in testing, being fixed. No launch yet.
Learnings
A requirement can be "non-negotiable" on paper and still fail in testing, hallucination control looked solid until a basic FAQ test proved otherwise.
Writing short is harder than writing thorough. The character-limit problem forced real decisions about what the bot actually needs to know first.
Writing short is harder than writing thorough. The character-limit problem forced real decisions about what the bot actually needs to know first
Next Steps
Retest against the same FAQ scenarios that surfaced the hallucination, confirm the fix actually holds.
Get real prospect conversations in front of the bot instead of only internal test scenarios.
Bring the business-pricing question back to Carmelina directly, rather than leaving the bot to improvise an answer.
Personal Growth
Strength in building AI knowledge base
Fine tuning and training output and hallucinations
Spelling out the invisible parts of a conversation (rules for outputs, tone, when to stay quite, when to hand off, what a good answer looks like, etc.).
