Readable transcripts
HTML and text views organize dialogue into readable turns. A local comparison checks preservation of speech text during formatting, ignoring whitespace. Uncertain speaker attribution stays labelled.
AI Meeting-to-Action
From discussion recordings to reviewable documents and action plans.
A locally hosted workflow that combines Python, n8n and the OpenAI API to turn audio or text into a readable transcript, a Word summary, and optional editable PowerPoint and Excel deliverables.
Explore a simulated workflow with synthetic data and downloadable sample documents, or watch a 3-minute 53-second Full HD walkthrough explaining the actual local n8n workflow through native screen captures and Indonesian narration. No API calls are made by this public demo.
01 / Problem
A meeting ends. The follow-up work begins.
Reviewing recordings, preparing summaries, creating presentation material and assigning follow-up tasks can involve repeated manual work. This project brings those steps into a resumable workflow while retaining links to transcript evidence and making uncertain results visible.
02 / Workflow
One source, several useful outputs
01 · InputAudio or transcriptUpload a recording, upload text, or reuse a saved transcript.
02 · ExtractTranscript + evidenceTranscribe audio and extract facts with source segment references.
03 · ReviewSummary + quality checksCheck claims against transcript context and flag results that need review.
04 · DeliverWord, PPTX, ExcelChoose summary only or the presentation and planning workflow.
03 / What I built
Automation with checkpoints
Flexible deliverables
Summary-only mode produces a Word report without presentation planning. The extended mode adds editable slides and an Excel action plan. AI recommendations are distinguished from recorded decisions.
Resumable orchestration
n8n coordinates stages through a local FastAPI service. Saved checkpoints and response caching allow completed work to be reused when the workflow resumes.
Cost-aware execution
Call limits, persistent caching, disabled automatic network retries and local rendering checks help control unnecessary paid requests. These controls are not a guarantee of a fixed total bill.
04 / Architecture
Local coordination, API-based AI processing
Upload form → n8n orchestration → Python/FastAPI worker → transcription and evidence extraction → summary and quality review → optional recommendation planning → local document rendering.
The application and stored jobs run locally. The current transcription path sends audio to OpenAI, and AI analysis sends transcript text to the API. Document rendering runs locally with python-docx and PptxGenJS; Excel export is part of the planning workflow. This is a controlled workflow with AI stages.
05 / Validation
What has been checked
The recorded synthetic-text demo exercised the real n8n workflow. Its initial outputs required review. All findings used for the final revision were addressed in a documented editorial review, with revised PowerPoint and Excel exports. The latest AI plan audit did not pass the preceding version; no new AI-audit pass is claimed. The recorded n8n execution and its original outcome remain preserved. The narrated full-screen video explains the distinction.
Project records document automated checks for output modes, transcript formatting, source references, cost controls, checkpoint reuse and workflow contracts. Many AI-path tests use simulated responses. A recorded five-minute transcription API trial establishes that an API path was exercised, but does not establish accuracy or end-to-end performance. No accuracy percentage, time-saving percentage or production SLA is claimed.
06 / Limitations
Review remains part of the process
- Speech recognition and speaker labels may be wrong; formatting does not verify them against audio.
- AI review checks consistency with transcript evidence and can still miss errors.
- Recommendations are proposals for human review, not decisions made during the meeting.
- The host computer and required services must stay available; API stages require internet and paid API access.
- This page is a project case study. Source code, private recordings and client documents are not published here.