AI2025
zbardh.al
AI meeting transcription & document automation for municipalities.
zbardh.al is an AI-powered meeting-transcription and document-automation platform I build at Nullius In Verba for Albanian municipalities (bashki). Users upload meeting audio, which is automatically transcribed with speaker diarization, summarized, answered against custom questions, and exported into formatted official documents.
The hard part is an end-to-end audio→document pipeline that stays accurate and correctable, plus per-user prompt templates that produce consistent official minutes.
Architecture
- Pipeline: in-memory FFmpeg conversion → S3 object storage → speech-to-text with diarization & word timestamps via async webhook callbacks.
- AI: LLM summarization and configurable per-user question-answering (system prompt, format, JSON schema), plus a context-aware chat assistant.
- Documents: templated DOCX/PDF generation (docxtemplater, LibreOffice, Puppeteer) for official minutes; segment-level transcript correction.
- Stack: NestJS 11 + Prisma + PostgreSQL, Next.js 15 / React 19 dashboard with a wavesurfer.js waveform player; Turborepo + pnpm; DigitalOcean Spaces.
Tech stack
NestJSPrismaPostgreSQLFFmpegAWS S3 SDKSpeech-to-TextDiarizationLLM SummariesQ&ANext.js 15React 19wavesurfer.jsTanStack QueryTurborepopnpmDockerDO Spaces