Sandwich — Voice-Driven Development Workflow

Turning everyday downtime into productive dev time.

Comment

🥪 Sandwich — Voice-Driven Development Workflow

Ship features while you’re grabbing a sandwich—Async dev, powered by Real-time Voice

Inspiration

We’ve all been there — the deadline looms, but you’re away from your keyboard. Whether it’s commuting, grabbing lunch, or running errands, those lost minutes add up.
What if development didn’t have to pause when you did?
That’s the idea behind Sandwich: turning everyday downtime into productive dev time.

What it does

Sandwich lets developers call in and give natural-language instructions to an AI voice agent, which then:

  1. Parses the request with Claude (Anthropic)
  2. Translate it into concrete dev tasks
  3. Gets the Human in the loop — the system calls the dev back to:
    • Confirm or refine the intent
    • Collect additional input over the phone
    • Keep development moving with real-time feedback
  4. Updates the codebase through code-gen agents
  5. Sends status updates and summaries via call or email

All async. All while you're away from the keyboard.

How we built it

  • Voice Agentic Framework: We used Vapi to capture full conversational voice input and convert it into structured prompts for downstream processing.
    • Orchestration (Claude API): Claude served as the central brain — orchestrating task flows, interpreting developer intent, and managing context across multiple instructions.
    • Code Generation (Claude): We used Claude not just for orchestration, but also as the codegen agent — generating and editing code snippets based on natural language prompts.
    • Frontend (Flask UI): A lightweight Flask-based web interface allowed us to monitor agent status and trigger voice sessions manually for testing and demo purposes.
    • Routing & Integration: The system links Vapi (voice), Claude (intent + code), and our backend logic into a seamless async pipeline.

Challenges we ran into

  • Managing agentic memory across sessions to retain context.
  • Interpreting ambiguous prompts (e.g. “make the button pop more”) into concrete actions.
  • Building a feedback loop that feels intuitive and safe — voice dev has to earn trust.
  • Handling fallback gracefully when the voice pipeline fails. ## Accomplishments that we're proud of

What we learned

  • Voice isn’t just a gimmick — it’s a legitimate interface for dev workflows.
  • Agent orchestration is hard, especially when state and timing are asynchronous.
  • Developers are open to non-traditional UIs if they reduce friction and speed up shipping.

What's next for Sandwich

  • Add live preview links when changes are made, sent via text or email.
  • Allow multi-turn voice convos for more complex flows.
  • Expand to support standup reporting, issue triage, and PR reviews — all by voice.
  • Integrate directly into tools like Slack, Linear, and GitHub Actions.

Built With

Share this project:

Updates