Skip to content
Amogh Suman

Reviewing 600,000 conversations for safety, in 24 languages

A multi-layer AI review of user conversations on a large consumer platform, combining rules, language models, voice analysis and trained human reviewers.

Sector
Consumer platforms and financial services
Scope
600K+ sessions across 24 languages, text and voice
My role
Architect and delivery lead

The problem

A large consumer platform needed to know how its conversations held up on safety, across hundreds of thousands of sessions, in two dozen languages, in text and in voice.

Reading everything by hand was impossible, and keyword filters miss most of what matters, especially in voice and in regional languages. The review also had to stand up to scrutiny, so the method mattered as much as the findings.

The approach

  1. Rules take the clear cases

    Pattern matching catches the obvious cases cheaply and keeps them out of the expensive stages.

  2. Language models classify the rest

    Each remaining conversation is classified against a written safety policy, with structured outputs so that results can be counted and checked.

  3. Voice gets its own analysis

    Acoustic analysis of voice sessions picks up what a transcript alone misses.

  4. Trained reviewers check a stratified sample

    A team of more than 10 reviewers checks a sample drawn across languages and categories, so the findings extend to the full set with a known level of confidence.

  5. Keep model costs in check

    When language model costs ran over plan, reworking the prompts and moving to structured outputs brought them back in line.

Who it serves

  • Trust and safety teams

    Where the risk sits, by category, language and channel.

  • Leadership and boards

    An independent view, with a method they can explain and defend.

  • Financial services firms

    The same pipeline, applied to customer conversations with banks and insurers.

The outcome

More than 600,000 sessions reviewed across 24 languages, with findings that rest on a documented sampling method.

The pipeline now runs as a repeatable review, and has since been used for several financial services clients.