ExpertAI

Ask a legal, financial, or medical question and get a domain-specialized AI answer in seconds — backed by licensed professionals when you need more than advice.

Comment

Inspiration Legal, financial, and medical information is expensive, slow, and rarely available at the moment it matters most. A tenant facing a lease at 11pm, a freelancer filing taxes for the first time, or a patient preparing for an appointment has no affordable on-ramp — and the professionals who could help are drowning in repetitive intake work instead of applying judgment.

We built ExpertAI to change that division of labor: AI handles the routine knowledge work around the clock, and qualified humans hold the judgment for the cases that actually need it.

What it does ExpertAI is an AI-operated gateway to professional services. A user asks a question in plain language, and a team of Gemini agents takes over:

TriageAgent classifies the domain (legal, financial, or medical), estimates complexity, and flags whether human judgment is required. Specialist agents (Legal, Financial, Medical) draft an educational response with built-in disclaimers and refusal rules. FollowUpAgent recommends next steps; EscalationAgent packages high-risk matters into an intake brief and routes them to a qualified professional. Every decision — agent, action, decision, confidence, latency — is persisted as an auditable execution trail, viewable in the product. AI never answers high-risk matters alone, and it never impersonates a human. Safety is a boundary, not a disclaimer.

How we built it We started with one principle: accountability. So agents have no tools, no database access, and no ability to overstep — a FastAPI router validates and persists every business action. We built the orchestration and Gemini agent layer first, then the Next.js dashboard, query workspace, operations view, and professional portal, then the Stripe subscription flow, and finally the email and security layer.

On the platform side: FastAPI + SQLAlchemy on managed Postgres, Gemini via google-genai, live Stripe billing ($19/mo Individual, $99/mo Professional) with verified webhooks, Resend/SMTP for transactional email, and rate limiting, security headers, and input sanitization throughout. We shipped continuously, deployed on Render, and let production metrics tell us what mattered.

Challenges we faced Making AI accountable. Free-roaming agents are a demo; accountable agents are a product. We designed a structured execution log so every decision can be audited and replayed. Prompt injection and safety. We treat all user input as data and gate high-risk requests behind human escalation rather than the model's judgment alone. Real billing reliability. Our first live payment was almost lost to a failed webhook delivery while the API was crash-looping. We fixed the environment configuration and built Stripe↔database reconciliation so a successful payment is never stuck behind a delayed webhook — the plan self-heals on the next login. Building for real users. Recruiting one real paying customer, keeping the AI completion rate at 86%, and disclosing a true P&L with all acquisition spend taught us more than any metric dashboard could. What we learned AI and professionals are not a zero-sum trade. Every escalated case is a paid referral for a lawyer, CPA, or clinician; every answered question is access for someone who previously had none. The future of professional services is not automation replacing people — it's automation doing the intake, and people doing the judgment.

$$ \text{AI completion rate} = \frac{\text{queries resolved autonomously}}{\text{total queries}} = 86% $$

Built With

Share this project:

Updates