Blog Post Plan: Distill - Your Local AI-Powered Meeting Assistant

Issue: PRO-144
Status: Plan Created - Awaiting Approval
Date: 2026-04-06

Target Audience

• Privacy-conscious professionals and researchers
• Tech-savvy users who want local-first AI tools
• People tired of cloud-based transcription services
• Open-source enthusiasts

Blog Post Structure

1. Hook: The Meeting Recording Problem

• We record countless meetings (Zoom, in-person, phone calls)
• Traditional solutions: Otter.ai, Rev.com, Google Meet transcription
• The catch: Your sensitive conversations go to the cloud
• Privacy concerns for researchers, academics, corporate users

2. Enter Distill: The Local-First Solution

• CLI tool that transcribes, diarizes, and summarizes audio/video
• Key differentiator: Everything runs locally except optional LLM call (you control it)
• Perfect for confidential project discussions, academic research, sensitive meetings
• Inspired by: https://github.com/pasrom/meeting-transcriber
• What’s different: Turns it into a flexible CLI, recording app-agnostic process
• The magic: Record from ANY app, use Syncthing’s robust syncing to drop audio in a folder, machine listens for new files, processes automatically, outputs protocols to archive folder or further LLM processing
• Real integration: Hermes agent grabs protocols and merges them with a knowledge base it maintains

3. How It Works

audio/video file
      ↓
   ffmpeg          → convert to 16kHz mono WAV
      ↓
   faster-whisper  → transcription (Linux/CUDA) or mlx-whisper (macOS)
      ↓
   pyannote.audio  → speaker diarization (who spoke when)
      ↓
   speaker DB      → match voices against known speakers
      ↓
   local LLM       → protocol generation (summary, action items, decisions)
      ↓
   .txt + .md      → clean, structured output

4. Features That Make It Special

• Speaker Persistence: Once you label a speaker (“Alice”), distill recognizes them across all future recordings using voice embeddings
• Agent-Friendly Design: Built for automation — works seamlessly with Syncthing for automatic processing
• Flexible LLM Backend: Works with Ollama, llama.cpp, LM Studio, vLLM — you choose your model
• Small Model Friendly: Designed to work with models like Gemma 3 4B, making it laptop-friendly
• Multi-Format Support: Handles .wav, .mp3, .m4a, .mp4, .mkv, .webm, .ogg, and more

5. Real-World Use Case: Professor Workflow

• Record meetings on phone/watch/laptop
• Syncthing syncs recordings to desktop
• Distill automatically processes new files
• Telegram bot helps identify speakers via audio samples
• Output: Professional meeting protocols in Markdown
• All local, all private, all automated

6. Installation & Quick Start

git clone https://github.com/ilyasst/distill
cd distill
python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[linux]"  # or ".[macos]"

# Process a meeting
distill process meeting.wav \
  --llm-endpoint http://localhost:8080/v1 \
  --llm-model gemma3:4b \
  --output ./results

7. What You Get

• Diarized transcript (who said what, with timestamps)
• Structured protocol with:
  • Summary
  • Participants
  • Topics Discussed
  • Decisions Made
  • Action Items
  • Open Questions

8. Privacy & Data

• No audio ever leaves your machine
• Speaker database stored locally at ~/.distill/speakers.json
• You control the LLM endpoint
• Open source: Audit the code yourself

9. Limitations

• Requires some setup (not a one-click install)
• Diarization needs HuggingFace token (free, but required)
• Processing speed depends on hardware
• Small Whisper models may have lower accuracy for noisy recordings

10. Call to Action

• Try it: GitHub repo link (https://github.com/ilyasst/distill)
• Contribute: Open source, welcomes PRs
• Share your workflow: How are you using distill?

Visual Elements to Include

• Architecture diagram (ASCII or image)
• Sample input/output comparison
• Screenshot of a generated protocol
• Terminal usage examples with colored output

SEO Keywords

• local transcription
• private AI
• meeting assistant
• speaker diarization
• open source transcription
• offline AI tools
• CLI productivity

Related Blog Posts to Link

• Whisper model comparisons
• Local LLM setup guides
• Privacy-first AI tools

APPROVAL STATUS

• Plan created: 2026-04-06
• Plan approved: YES (2026-04-06)
• Feedback incorporated: Added pasrom/meeting-transcriber inspiration, CLI/app-agnostic emphasis, Syncthing workflow, Hermes KB integration
• Next step: Write full blog post and push as hidden post to _posts/

Research Notes

Repository explored: ~/Repositories/ilyasst/distill/

Key files analyzed:

• README.md: Comprehensive documentation of the tool
• distill.py: Main Python implementation
• distill-server.py: Server component for automation
• jobs/: Contains processed meeting examples with protocols

Key technical details discovered:

• Uses cosine similarity for speaker matching (threshold: 0.82)
• Stores up to 5 embeddings per speaker
• Speaker IDs format: SPK_XXX (e.g., SPK_001)
• Cache directory for diarization results: ~/.distill/cache/
• Speaker database: ~/.distill/speakers.json

Sample output format:

• Meeting protocols include: Summary, Participants, Topics, Decisions, Action Items, Open Questions
• Full transcript with speaker labels and timestamps
• Both .txt (diarized transcript) and .md (structured protocol) outputs