Use cases · Voice of customer

How it works

Every review, ticket and survey answer goes first to the small model, which labels the ones it is confident about — the clear majority — and hands them straight to your spreadsheet. What it is unsure about, it sets aside instead of guessing: sarcasm, backhanded praise, a complaint in a minor EU language. That residue is queued and read overnight by the large model on ml2, the machine with the bigger card — the model is far too big to sit on a graphics card, so its bulk rests in slower system memory and only a small part of it works on any one word, which is why it runs as an overnight batch and not while anyone waits. The whole point of the two tiers is cost: the expensive model only ever sees the hard residue, which is what makes running it affordable at all.

Data you need

The same feedback exports as the rest of this group — reviews, tickets, survey answers as text — plus your label codebook and a rule for what counts as "unsure", usually the small model's own low-confidence items with whole categories like a named minor language added by hand. It fails when there is more residue than a night can hold: a vague codebook, or a pile of feedback that is mostly in a language the small model reads poorly, sends most of the pile to the overnight queue until it no longer finishes by morning and the saving disappears. Feedback that arrives only as screenshots or scanned cards has no text to read and is skipped by both tiers.

What to expect

Where the small model is genuinely confident its labels are as good as running it alone, and the overnight pass lifts quality on exactly the items a lone small model would have quietly mislabelled. The routing is the weak point: the small model does not always know when it is wrong, so a sarcastic review it misreads with confidence gets a firm wrong label and never reaches the large model at all. A flat "well, it eventually turned up" is filed as praise by the first tier, the large model never sees it, and no one queries it — the error is invisible precisely because the small model was sure. And because the hard items wait for the overnight run, nothing here happens while you wait: an urgent safety complaint hiding in the ambiguous pile is not seen until the next morning, so this must never be your safety or legal routing path, which stays live and separate.

Where people stay involved

A person owns the codebook and the confidence threshold that decides what escalates — set it too tight and the large model drowns, too loose and the hard items sail through mislabelled. Spot-check both piles, not only the escalated one, because the failures that cost you are the items the small model was wrongly sure about. Never let this pattern carry safety, legal or money decisions: its whole design holds the hard cases back to overnight, which is the opposite of what those need.

• Turn a quarter's worth of product reviews into a short themes report

  Turns exported product reviews into a short plain-language themes report, with every claim linked back to actual review text.

• Tag every review and ticket with your own fixed label list, so you can finally count things

  Applies your own fixed label list to every review, ticket or survey answer, turning free text into a column you can pivot and chart.

• Summarise what your NPS promoters and detractors actually wrote

  Summarises the free-text answers behind your NPS score, band by band — what promoters praise and detractors cite, backed by direct quotes.

• A Monday-morning digest of what customers complained about last week

  A one-page weekly digest of support tickets — top issues, movement against last week, anything unusual — in the inbox before Monday stand-up.