Every review, in one place

All the reviews with text from the twelve-week window. Filter by theme, rating or a word, and read them as the users wrote them. Display names were dropped at ingest; ids are hashed.

Five themes, and the words that mark them

A review goes to its strongest theme only, so nothing is counted twice. Reviews that match nothing are left out rather than forced into the nearest bucket. The distinguishing words are the data's own, picked by log-odds against the rest of the corpus.

Which charge is misunderstood, not merely disliked

A review counts as confused when it names a charge and says it was not expected: hidden, extra, undisclosed, bina bataye, pata nahi. Volume alone would always pick the best-known charge, so the score multiplies how often a charge is raised by how often it is misread.

For the product team, every Monday. The top themes, the quotes behind them, one observation about what is going wrong, and three things to do about it. Capped at 250 words, so it gets read.

For support to paste when a user asks why something was deducted. Facts and figures from official pages only, in the order a confused user asks them. No reassurance, no recommendation, and a date so it cannot go stale silently.

Nothing is logged until you approve it

Two MCP tools are waiting. The server (the same src/mcp-core.js that mcp/server.js runs over stdio, hosted here in the page) has refused both calls once and returned exactly what it would write, with a single-use token bound to that payload. Read the payloads, then decide.