Boswell: AI Research Interviewer

An open-source AI interviewer that conducts research-informed voice interviews autonomously and outputs structured transcripts and insights.

Overview

Boswell is designed for researchers, authors, and journalists who need to conduct substantive interviews at scale without losing the human touch. You provide a topic and research materials, Boswell generates tailored interview questions, then conducts a real-time voice interview using AI.

How it works:

1. Provide a topic and research materials (PDFs, documents, URLs)
2. Boswell generates informed interview questions using Claude
3. Start a voice interview session - Boswell creates a video room you share with your guest
4. The AI conducts a dynamic, conversational interview following interesting threads
5. Export clean transcripts and AI-generated insights

Key Features:

• Research-informed question generation from documents and URLs
• Real-time voice interviews with natural conversation flow
• Dynamic follow-up questions that pursue interesting threads
• Pause & Resume - Stop interviews and continue later with full context preserved
• Dynamic Speed Control - Guests can ask the bot to speak faster or slower
• Strike from Record - Guests can say "forget that" to remove content from transcript
• Privacy First - Anonymous by default, transcript emailed to guest after
• Immediate acknowledgments ("Mm-hmm", "I see") for natural conversation pacing
• Automatic transcript capture and insight extraction
• Low-latency voice synthesis (~500ms response time)

Architecture

┌─────────────────────────────────────────────────────────────┐
│                        BOSWELL CLI                          │
│     boswell create | start | status | export | list         │
└─────────────────────────────────────────────────────────────┘
                              │
┌─────────────────────────────────────────────────────────────┐
│                     VOICE PIPELINE                          │
│  ┌──────────┐   ┌────────┐   ┌────────┐   ┌────────────┐   │
│  │ Daily.co │──▶│Deepgram│──▶│ Claude │──▶│ ElevenLabs │   │
│  │Transport │   │  STT   │   │  LLM   │   │    TTS     │   │
│  └──────────┘   └────────┘   └────────┘   └────────────┘   │
│       │              │            │             │           │
│       ▼              ▼            ▼             ▼           │
│    [Audio In]   [Transcript]  [Response]   [Audio Out]     │
└─────────────────────────────────────────────────────────────┘
                              │
┌─────────────────────────────────────────────────────────────┐
│                    EXTERNAL SERVICES                        │
│  • Daily.co    - WebRTC video rooms                         │
│  • Deepgram    - Speech-to-text (Nova-2)                    │
│  • Claude      - Conversation intelligence (Sonnet 4)       │
│  • ElevenLabs  - Text-to-speech (Turbo v2)                  │
└─────────────────────────────────────────────────────────────┘

Quick Start

Prerequisites

• Python 3.11+
• API keys for: Claude (Anthropic), Daily.co, Deepgram, ElevenLabs

Installation

# Clone the repository
git clone https://github.com/noahraford/boswell
cd boswell

# Create virtual environment
python3 -m venv venv
source venv/bin/activate

# Install with voice dependencies
pip install -e ".[voice]"

# Initialize configuration
boswell init

Your First Interview

# 1. Create an interview with research materials
boswell create --topic "AI Safety Research" --docs ./research/paper.pdf

# 2. Start the voice bot
boswell start int_abc123

# 3. Share the Daily.co room URL with your guest
#    The bot will greet them and conduct the interview

# 4. Export transcript and insights when done
boswell export int_abc123

CLI Commands

boswell init

Initialize Boswell with your API keys.

boswell init

Prompts for:

• Claude API Key (required) - For question generation and conversation
• Deepgram API Key (required) - For speech-to-text
• ElevenLabs API Key (required) - For text-to-speech
• Daily.co API Key (required) - For video rooms

boswell create

Create a new interview session with research materials.

boswell create --topic "Future of Work" --docs ./research.pdf,./notes.md --urls https://guest-bio.com

Options:

• --topic, -t (required): Interview topic
• --docs, -d: Comma-separated paths to research documents (PDF, TXT, MD)
• --urls, -u: Comma-separated URLs to scrape for research

What happens:

1. Ingests and processes research materials
2. Generates tailored interview questions using Claude
3. Creates an interview record ready for voice session

boswell start

Start a voice interview bot.

boswell start int_abc123

What happens:

1. Creates a Daily.co video room
2. Starts the Pipecat voice pipeline
3. Bot joins and waits for guest
4. Displays shareable room URL
5. Conducts interview when guest joins
6. Press Ctrl+C to pause - saves conversation context for later
7. Saves transcript when complete

Output:

Starting voice interview: int_abc123
Topic: AI Safety Research
Questions: 12

Daily.co room created!
==================================================

Send this link to your guest:
  https://yourname.daily.co/boswell-int_abc123

==================================================

Bot is joining the room...
Press Ctrl+C to end the interview.

boswell resume

Resume a paused interview with full context preserved.

boswell resume int_abc123

What happens:

1. Creates a new Daily.co room
2. Loads conversation history from the paused session
3. Bot welcomes guest back and continues where you left off
4. All previous context is preserved

Tip: Pause/resume is useful for:

• Breaking long interviews into multiple sessions
• Handling technical difficulties
• Giving guests time to gather thoughts

boswell status

Check interview status.

boswell status int_abc123

Statuses:

• pending - Created, not started
• waiting - Bot in room, awaiting guest
• in_progress - Interview happening
• paused - Interview paused, can be resumed
• complete - Interview finished
• error - Something went wrong

boswell export

Export transcript and insights.

boswell export int_abc123 --output ./interviews/

Options:

• --output, -o: Output directory
• --transcript, -t: Path to external transcript JSON (optional)

Outputs:

• transcript.md - Clean, formatted interview transcript
• insights.md - Key themes, notable quotes, and summary

boswell list

List all interviews.

boswell list

boswell retry

Retry a failed or no-show interview.

boswell retry int_abc123

Configuration

Config File

Located at ~/.boswell/config.json:

{
  "claude_api_key": "sk-ant-...",
  "deepgram_api_key": "...",
  "elevenlabs_api_key": "...",
  "daily_api_key": "...",
  "default_target_time": 30,
  "default_max_time": 45
}

Environment Variables

Alternatively, use a .env file:

CLAUDE_API_KEY=sk-ant-...
DEEPGRAM_API_KEY=...
ELEVENLABS_API_KEY=...
DAILY_API_KEY=...

Interview Flow

┌─────────────────┐
│  boswell create │  ← Provide topic + research
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│  Questions      │  ← Claude generates tailored questions
│  Generated      │
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│  boswell start  │  ← Creates Daily.co room
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│  Share URL      │  ← Guest joins video room
│  with Guest     │
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│  Voice          │  ← Real-time conversation
│  Interview      │     Deepgram → Claude → ElevenLabs
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│  boswell export │  ← transcript.md + insights.md
└─────────────────┘

The Interview Experience

When your guest joins the Daily.co room:

1. Greeting - Boswell introduces itself, explains:
   • The interview topic and expected duration
   • That the interview is anonymous (not attributed unless they want)
   • They'll receive a full transcript via email after
   • They can say "nevermind, forget that" to strike anything from the record
   • They can pause, stop, or ask for repeats anytime
2. Core Questions - Works through research-informed questions (one at a time, no sub-questions)
3. Acknowledgments - Immediate "Mm-hmm", "I see" responses to show it's listening
4. Follow-ups - Pursues interesting threads that emerge naturally
5. Check-ins - Periodically asks about time/comfort
6. Wrap-up - Thanks guest, asks one time for final thoughts, then goodbye

The AI interviewer is designed to be:

• Warm and curious - Like an NPR interviewer
• Research-informed - Questions reflect the provided materials
• Adaptive - Follows interesting threads rather than rigid scripts
• Respectful - Moves on gracefully if guest is uncomfortable
• Responsive - Guests can say "slow down" or "speed up" to adjust speech rate
• Privacy-conscious - Anonymous by default, with ability to strike content from record

Output Format

transcript.md

---
interview_id: int_abc123
topic: AI Safety Research
date: 2024-01-22
duration: 28 minutes
---

# Interview Transcript

**Boswell:** Before we get into the specifics, I want to understand
why you first got interested in AI safety. What was the moment or
realization that drew you to this work?

**Guest:** It actually started during my PhD in physics. I was
working on complex systems and started seeing parallels...

**Boswell:** That's fascinating - the connection to complex systems.
Can you say more about what parallels you were seeing?

insights.md

# Interview Insights

## Key Themes

### 1. From Physics to AI Safety
The guest's background in complex systems shaped their approach...
> "The problems felt similar - you're reasoning about systems
> you can't fully observe"

### 2. The Coordination Challenge
...

## Notable Quotes

> "We're building systems that will be smarter than us, and we
> don't have a good theory of how to keep them aligned with
> human values"

## Summary

[AI-generated 2-3 paragraph overview]

Project Structure

boswell/
├── src/boswell/
│   ├── cli.py           # Command-line interface
│   ├── config.py        # Configuration management
│   ├── interview.py     # Interview model & lifecycle
│   ├── ingestion.py     # Research processing
│   ├── output.py        # Transcript & insights export
│   └── voice/
│       ├── bot.py           # Interview bot lifecycle
│       ├── pipeline.py      # Pipecat voice pipeline
│       ├── prompts.py       # System prompts for Claude
│       ├── transcript.py    # Transcript capture
│       ├── acknowledgment.py # Filler words ("Mm-hmm")
│       ├── speed_control.py  # Dynamic speech rate control
│       └── strike_control.py # "Forget that" content removal
├── tests/
├── docs/
└── pyproject.toml

Development

# Install dev dependencies
pip install -e ".[dev,voice]"

# Run tests
pytest tests/ -v

# Lint
ruff check src/ tests/

Service Costs

Service	Purpose	Pricing
Daily.co	Video rooms	10k free mins/month, then $0.004/min
Deepgram	Speech-to-text	$0.0043/min (Nova-2)
ElevenLabs	Text-to-speech	$0.30/1k chars (Turbo)
Claude	LLM	$3/$15 per 1M tokens (Sonnet)

Estimated cost per 30-min interview: ~$2-4

Roadmap

• Research ingestion and question generation
• Real-time voice interviews via Daily.co + Pipecat
• Transcript capture
• Basic export (transcript.md)
• Pause & resume interviews with context preservation
• Dynamic speech speed control (guest can request faster/slower)
• Immediate acknowledgments for natural conversation flow
• Strike from record ("forget that" removes content from transcript)
• Insights generation with quotes
• Cloud deployment option
• Recording and playback
• Multi-language support

License

MIT

Acknowledgments

Built with:

• Pipecat - Voice AI framework
• Daily.co - WebRTC infrastructure
• Deepgram - Speech-to-text
• ElevenLabs - Text-to-speech
• Anthropic Claude - LLM

---

Named after James Boswell, the 18th-century biographer famous for his detailed, conversational biography of Samuel Johnson.