Foresight

Turn recorded conversations into structured insights. Automatically.

Drop an audio file in a folder. Get back a searchable transcript and an AI-extracted summary with key themes, quotes, and follow-up questions.

100% local. No API costs. No data leaves your machine.

---

The Problem

You record important conversations—interviews, user research calls, sales demos, meetings. Then the audio sits there because:

• Transcription services cost money and take time
• Reading a full transcript is tedious
• The insights you need are buried in 45 minutes of audio
• Organizing files manually is a chore

The Solution

Audio file drops in folder
         ↓
   [3-5 minutes]
         ↓
Transcript.md + Insights.md
         ↓
Original archived automatically

What you get:

Output	Contents
Transcript	Full searchable text, YAML metadata, word count
Insights	Pain points, objections, decision criteria, key quotes, follow-up questions, executive summary

---

Use Cases

• User Research — Extract patterns from customer interviews
• Sales Calls — Capture objections and buying signals
• Podcasters — Generate show notes and quotable moments
• Journalists — Transcribe interviews with structured highlights
• Founders — Process investor/customer conversations at scale
• Consultants — Document client discovery sessions
• Academics — Transcribe and analyze qualitative research

---

Install

pip install foresight-transcribe

That's it. Or install from source:

git clone https://github.com/dnoma/foresight.git
cd foresight
pip install -e .

Setup Ollama (required for insights)

# Install Ollama
brew install ollama   # macOS
# or: curl -fsSL https://ollama.ai/install.sh | sh  # Linux

# Start and pull model (~4GB)
ollama serve &
ollama pull mistral

Verify installation

foresight --test

---

Usage

Watch Mode (recommended)

Leave it running. Drop files in the folder, they process automatically.

foresight --watch

Single File

Process one recording right now.

foresight --file meeting.m4a

Batch Mode

Have a backlog? Process everything at once.

foresight --batch

---

What the Output Looks Like

Transcript

---
filename: customer-call-jan-11.m4a
duration: 34:22
word_count: 4521
model: whisper-medium
---

# Transcript: customer-call-jan-11

The full conversation, searchable and quotable...

Insights

## PAIN POINTS
- Manual invoice processing takes 3 days each month
- No visibility into supplier risk until problems occur

## DECISION CRITERIA
- Must integrate with existing SAP system
- Needs to show ROI within 90 days

## KEY QUOTES
- "We've been burned twice by suppliers going bankrupt with no warning"

## FOLLOW-UP QUESTIONS
- What's the current approval workflow for new suppliers?
- Who else is involved in the vendor selection process?

## EXECUTIVE SUMMARY
[2-3 sentence summary of the entire conversation]

---

Directory Structure

Auto-created on first run:

~/klavis-interviews/
├── raw/           ← Drop recordings here
├── transcripts/   ← Whisper output
├── insights/      ← LLM-extracted analysis
├── archive/       ← Processed originals
└── processing.log

---

Configuration

Edit the config in interview_processor.py (or after install: ~/.local/lib/python*/site-packages/interview_processor.py):

WHISPER_MODEL = "medium"    # tiny|base|small|medium|large
OLLAMA_MODEL = "mistral"    # or phi3, llama3, etc.
MIN_FILE_SIZE = 100000      # Skip files under 100KB

Customize the extraction prompt

The EXTRACTION_PROMPT variable controls what the LLM extracts. Modify it for your use case:

# For sales calls
EXTRACTION_PROMPT = """Extract: objections raised, competitor mentions,
next steps agreed, budget signals..."""

# For user research
EXTRACTION_PROMPT = """Extract: user goals, frustrations,
current workflow, feature requests..."""

---

Options

Flag	Description
-m, --model	Whisper model size (default: medium)
--skip-insights	Transcribe only, skip LLM extraction
--keep-original	Don't move to archive after processing
--dry-run	Preview what would be processed
--status	Show pending files and system health
--retry-failed	Retry previously failed files
--test	Verify installation works

---

System Requirements

• macOS (tested) or Linux
• ~6-7GB RAM during processing (models load/unload automatically)
• Python 3.9+
• ~5GB disk for models (one-time download)

Processing Time

Audio Length	Time (M1 Mac)
5 min	~45 sec
30 min	~3 min
60 min	~6 min

---

Why Local?

	Local (this tool)	Cloud APIs
Cost	Free	$0.006/min+
Privacy	Data stays on device	Uploaded to servers
Speed	No upload/download	Network dependent
Availability	Works offline	Requires internet

A 1-hour recording costs ~$0.36 on cloud transcription. Process 100 interviews and you've saved $36—plus your data never left your laptop.

---

Supported Formats

.m4a .mp3 .wav .mp4

Files under 100KB are skipped (filters out accidental recordings).

---

Troubleshooting

Ollama not running?

ollama serve        # Start the server
ollama ps           # Check if model is loaded

Transcription too slow?

foresight --file audio.m4a --model small

Check system status:

foresight --status

---

License

MIT

---

Contributing

PRs welcome. Ideas:

• Speaker diarization (who said what)
• Custom prompt templates via config file
• Web UI for reviewing insights
• Export to Notion/Obsidian
• Slack/Discord notifications when processing completes

---

Support

If this saved you time or money, consider giving it a star. It helps others discover the project.

Built by @dnoma — indie dev building tools that save time.