Skip to content
buzinov.tech, home page

Pipelines with LLMs

Call and document analysis with LLMs

Transcription, extraction and scoring that say how sure they are and never guess a missing value.

business rules checked on every invoice field
EN 16931
output validated against OASIS schemas
UBL 2.1

What you get

  • Calls. Recordings go through speech-to-text, then a language model scores each transcript against your QA checklist: greeting, needs discovery, objections, next step. Scores land in your CRM next to the deal.
  • Documents. Invoices, acts and contracts are read with OCR and a language model into structured fields. Every field is checked against rules before anything is saved.
  • A queue with retries and a log, so a failed file is visible and can be sent again.

How I build it

The model is one step of the pipeline, not the whole pipeline. Around it sit validation, confidence thresholds and a review queue for people. For invoices the rules come from EN 16931, and the UBL output is validated against the official OASIS schemas.

A value the model did not see stays empty. A pipeline that guesses looks fine in a demo and breaks accounting in the first month.

How this worked in real projects: sales call scoring, an outbound voice agent and invoice extraction checked against EN 16931.

When you need call or document analysis

  • A sales manager listens to a sample of calls and misses most of them.
  • Accountants retype invoices from PDF into 1C or an ERP.
  • You already tried a chat assistant and could not trust its answers.

Packages

Prices are fixed for the described scope. Larger scope gets its own estimate.

  • Basic

    $500

    • Transcription and scoring against your checklist
    • One batch of calls or documents
    • Results in a spreadsheet
  • Standard

    $1,200

    • Automated pipeline with queueing and retries
    • Results written to your CRM or a spreadsheet
    • Per-field confidence
  • Premium

    $2,500

    • Everything in Standard
    • Dashboard
    • Checklist tuned on your own data

Questions

Which speech and model services do you use?

Calls are transcribed with Yandex SpeechKit, and scoring and document extraction use language models through the OpenAI API. Both can be replaced with services you approve or run yourself.

What if the model is wrong?

Every field comes with a confidence value, and a value the model did not see stays empty instead of invented. Low-confidence results go to a person for review.

Can recordings stay inside our infrastructure?

Yes. The pipeline runs on your server; only the speech and model APIs you approve receive data, and they can be swapped for self-hosted ones.

    • Python
    • Yandex SpeechKit
    • OpenAI API
    • PostgreSQL

    Real estate agency, Kazakhstan

    Every sales call scored against the checklist

    Call recordings go through speech-to-text, then a language model scores each transcript against the QA checklist. The pipeline replaced a large share of manual call review.

    Read the project

    • Python
    • Speech-to-text
    • Speech synthesis
    • LLM

    Real estate agency, Kazakhstan

    Outbound voice agent for a sales team

    A voice agent that calls clients from a campaign list, keeps the conversation on script and records the outcome for the sales team.

    Read the project

  • EN 16931business rules on every field

    Open source

    Invoice extraction checked against EN 16931

    OCR and a language model read the invoice, business rules check each field, and the UBL 2.1 output is validated against the OASIS schemas.

    Read the project

A process still runs on copy and paste?

Describe it in a few paragraphs. You get the approach, the first risk and a fixed estimate within 24 hours.