Clerova: an AI agent that acts only with human approval by Harleen GhumanClerova: an AI agent that acts only with human approval by Harleen Ghuman

Clerova: an AI agent that acts only with human approval

Portfolio project
Harleen Ghuman
The problem Small businesses like property managers lose hours every day to repetitive emails: maintenance requests, tenant questions, and follow-ups. Generic AI tools are risky here because they can send the wrong thing to a real customer.
What I built Clerova is a live, multi-tenant SaaS that turns forwarded emails into structured requests. For each email it pulls out the key details, recommends a next step and drafts a reply. Nothing is sent or created until a person approves, edits or rejects it. Approved replies go out by email and stay threaded with the original request.
How it’s built
The AI reasoning (OpenAI structured outputs) is kept separate from the code that actually sends or creates anything, which is what makes human approval reliable
Email intake and outbound mail run on Postmark webhooks
Each customer’s data is kept fully separate (PostgreSQL row-level security), with per-organization AI quotas and rate limits
Auth with Supabase and Spring Security; Stripe billing with free trials and a customer portal
An AI test suite covers data extraction, drafted replies, and prompt-injection attempts hidden inside emails
Result I designed, built and launched it solo, end to end. It’s live in production with versions for property management and trucking.
Dashboard (demo data)
Dashboard (demo data)
Demo Data
Demo Data
Stack: Java, Spring Boot, React/TypeScript, PostgreSQL, OpenAI, deployed on Render with Cloudflare.

0 likes · 0 views · Nov 15, 2026

Designed, built and launched solo:an AI SaaS that turns business emails into structured requests and drafted replies with human approval before anything is sent

Tags