Reewu
An AI SaaS that helps local businesses collect authentic Google reviews. A QR scan pulls real business context into a prompt pipeline, so the generated review reads like the customer wrote it — not like generic AI text.
- Role
- Founding Eng
- Surface
- QR → Web
- AI Layer
- OpenAI + Claude
- Billing
- Stripe
The Friction
Businesses know reviews drive walk-in traffic, but the gap between a happy customer leaving the shop and writing a detailed review is enormous. Asking staff to chase reviews does not scale, and generic AI review generators produce text so obviously synthetic that it damages trust — and risks removal by the platform.
The Automation Flow
A QR code at the point of sale resolves to that specific business. Rather than prompting a model with "write a review", the platform assembles a prompt grounded in real, business-specific context — category, services actually offered, location, and the customer's selected sentiment — so the output is specific enough to be genuinely useful to the next reader.
System architecture
Hover or focus a stage to see why it’s built that way.
Grounding over cleverness
The quality difference between a generic prompt and one carrying real business context is larger than the difference between any two frontier models. Most of the engineering went into the context assembly layer, not model selection.
Provider-agnostic LLM interface
OpenAI and Claude sit behind a single internal interface. Pricing, latency and availability all move; being able to switch per-request without touching product code has repeatedly paid for itself.
MongoDB for business profiles
Business records are heterogeneous — a salon and a garage describe their services with completely different shapes. A document store avoided a migration every time a new vertical was onboarded.
Stack snapshot
- Next.js
- React
- Express
- MongoDB
- OpenAI API
- Claude API
- Stripe
- Tailwind CSS
Challenges & learnings
The authenticity barrier
Early output was fluent but hollow. Fixed by forcing every generation to cite at least one concrete, business-specific detail — a named service, the location, a specific offering — rather than letting the model free-associate.
Latency at the counter
A customer standing at a till will not wait. Streaming the response and pre-warming business context on QR resolve kept the perceived wait short enough that people finish the flow.
Prompts are code
Once prompts carry real business logic they need the same discipline as code — versioning, a fixture set of businesses, and a check that output still contains the grounding details before shipping a change.
One tap beats raw capability
Users never asked which model wrote the draft. Every conversion gain came from removing a step, not from improving generation quality.