Executive Summary: An LLM product built cost-first and governance-first, then audited against its own claims. SiftReviews is a Chrome extension that combines domain reputation signals with LLM-powered AI detection to evaluate digital authenticity, built on Gemini 3.6 Flash and FastAPI. V2 shipped rate limiting, hashed identifiers, and per-request cost tracking; V2.2 documents what a structured review found when those controls were tested against what the documentation claimed they did. V3 turns the same scrutiny on the model itself: an eval harness, the measurement failures it surfaced, and the two product fixes they forced. V3.1 corrected a pricing constant the instrument had been hiding, froze a regression suite, and shipped selected-text scanning. Published on the Chrome Web Store as v1.0.0.

Product · Design · Builder: Meghal Donde Pradhan · meghaldondepradhan.com

Independent project · August 2026

Quick Links:

🌐 Product Page | 🧩 Chrome Web Store | 💻 GitHub Repo | 📊 Live Admin Dashboard


1. The Problem

Reviews are where a shopper finds out what the product page will not say. Does it hold up after a month. Is it smaller than it looks. Did it arrive broken.

That check is getting harder. A lot of reviews now read the same way: even enthusiasm, a tidy structure, a recommendation at the end. Some of them are honest. Some are written to sell. Reading them does not tell you which.

The obvious question is whether a machine wrote it. That question is a dead end. A generated review can describe a real purchase accurately. A person can write three sentences that say nothing. Knowing who typed it does not tell a shopper whether to trust it, and it is the harder thing to find out.

The useful question is smaller. Which phrases in this review are doing the work, and does it say anything specific enough to act on. That is what SiftReviews answers. It shows the phrasing that makes a passage read as formulaic and leaves the judgment with the reader.


2. Who This Is For

These personas came out of field research on a single question: how does a shopper decide which reviews to believe. Guided tours, error analysis and extreme-user interviews, then the patterns clustered.

Persona What they need What gets in the way
Busy working parent A quick read on whether a review is worth weighting Too many reviews, too little time; decides on partial information
The researcher To trace a review back to its source and see what was discounted Integrity of the source matters more to them than the star average
Elderly browser A plain-language reason to trust or skip, without making an account Clutter and jargon raise suspicion and end in an abandoned cart
Newly independent teen buyer Confidence without wading through negative noise Anchors on the worst review; wants to avoid regret spending
Time-savvy professional shopper To filter authentic reviews quickly across sites Even skilled shoppers waste time cross-checking the same product

Three findings held across all of them:

That last one shaped the product more than anything else. It is why the extension shows quoted phrases from the passage instead of only a number.