What research showed
I read 32 public studies on retail investors to check whether the pattern held beyond one platform. It did. Three barriers kept coming back, and all of them had to do with feelings or knowledge.
of retail investors name fear of losing money as their main barrier (eToro, 2025).
of people don't understand investment fees (TenCap, 2024).
abandon financial app onboarding halfway through (Signicat, 2022).
Access was already there, and people still didn't feel ready. So I designed a mentor that makes the first steps feel safe.
Key decisions
- 01A mentor that never recommends
It explains concepts and scenarios and suggests the next question. It never says what to buy or sell, and every answer leaves the decision with the person.
- 02It starts with how you feel
The first question is how you feel about investing. The flow assumes anxiety, uses no-pressure language and reveals information in small steps.
- 03Two characters
Navi leads the explanations. Artemis, inspired by my own cat, notices hesitation and celebrates small wins. Together they make the AI feel like company.
- 04Every pattern traces to research
When to ask a follow-up, slow down or show a disclaimer came from studies on fear, drop-off and trust. That let me decide fast in 48 hours.
How the mentor evolved
I tested it the way a real beginner talks: changing direction mid-conversation and asking follow-ups like “and what if I changed my goal?”. Each failure became a design rule for the next version.
- V1One generalist agent
Fine for clean, one-off questions. Every message started from zero, so any follow-up broke trust.
- V2Specialized agents
Instruction, decision and clarity agents gave better answers, but only when people used the expected keywords.
- V3Router and memory
Intent detection plus short-term memory of goals, amounts and doubts. Follow-ups finally connected to the original plan.
- V4Calibrated mentor
Safety prompts, disclaimers and length limits made answers calmer, shorter and always educational.
Without structure and memory, a powerful model behaves like a static FAQ.
What I learned
The result is a working prototype that is ready to be validated with real investors and compliance teams.
In a conversational product, the dialogue is most of the design. Every line is a product decision, so I planned the empathy from the first message.
Guardrails can be part of the value. In a regulated market, an assistant that only educates earns more trust.
Knowing the backend widened my toolkit. Wiring agents, live market data and memory let me invent patterns like contextual follow-ups.
Try the prototype
Talk to Navi in your own words. Ask what a word means, look up a stock, or share a goal and then change your mind.
