All work

Local engineering project · ongoing development

Evidence-led call review

A local AI-assisted review application that keeps transcript evidence, provisional scores and human decisions visible.

  • Laravel
  • PHP
  • Livewire
  • Tailwind CSS
  • Whisper.cpp
  • Ollama
  • Queue Workers
Illustrative workflow · implementation study
  1. 01

    Recording

    Local upload and transcription

  2. 02

    Evidence

    Transcript-linked criteria

  3. 03

    Review

    Provisional scores stay visible

  4. 04

    Decision

    Human review with reasons

Case study

Overview

This local Laravel engineering project explores how to make recorded-call review inspectable. The work brings recording uploads, transcription, versioned evaluation criteria and reviewer decisions into one workflow.

Status: ongoing development. The case focuses on evidence and human review; model accuracy has not been independently evaluated.

Technical highlights

  • 01 Versioned rubrics preserve the context of each review.
  • 02 Transcript-linked evidence supports inspection of individual criteria.
  • 03 Final and provisional scores remain distinct from evidence coverage.
  • 04 Retry and recovery paths preserve processing context.
  • 05 Human overrides record reasons and guard against stale revisions.

Context

Problem

A score is hard to trust when the reviewer cannot see its evidence, the rubric version, or whether the entire recording was understood. Missing evidence should remain visible rather than quietly becoming a final judgment.

Solution

Queued processing connects Whisper.cpp transcription with Ollama evaluation. Criteria link back to transcript evidence, while a reviewer can record an override with a reason. Template versions keep the evaluation context explicit.

Technical architecture

Laravel and Livewire provide authenticated workflow screens. Upload batches, processing attempts and queued jobs separate intake from transcription and evaluation. Windowed evaluation and reconciliation make larger recordings reviewable. Output validators and the score calculator distinguish provisional scores, final scores and evidence coverage.

Challenges

Recording completeness, speaker attribution and uncertain criteria affect what a model can reasonably conclude. The implementation keeps unresolved criteria out of a final score and avoids whole-call omission judgments when completeness is uncertain. Model output still requires human review; no independent accuracy benchmark is claimed.

Working on something similar?

Tell me about the workflow, the existing system and what you want to change.

Discuss your project