Playbook

Inspiration

Sales outcomes are usually judged by transcripts and intuition, but the strongest signals often live in voice dynamics: hesitation, engagement drop-offs, tone shifts, and objection moments. We wanted to build an agent that doesn’t just talk—it learns from every call and continuously improves how it sells.

What it does

Playbook is a self-improving voice sales system: It conducts calls with ElevenLabs, captures call/transcript data in Supabase, analyzes emotional + behavioral voice patterns with Modulate Velma, triggers Airia to rewrite the sales playbook, applies that new playbook to future calls, and monitors performance trends in Lightdash.

How we built it

ElevenLabs powers the conversational voice agent with dynamic variables (strategy, opener, tone, etc.).

Supabase stores call records, analyses, and versioned playbooks. Modulate Velma processes call audio for engagement/tone/emotion/deception signals. Airia orchestrates a multi-node pipeline that fetches recent calls, reasons over patterns, generates a better playbook, and writes it back. Lightdash provides observability over version-by-version performance and call quality trends.

Challenges we ran into

Webhook reliability and event payload inconsistency for post-call audio/transcript. Distinguishing “triggered Airia” vs “Airia completed and persisted a new version.” Ensuring each call links to the intended active playbook version over time. Building UI clarity for rapid playbook evolution during live testing.

Accomplishments that we're proud of

Built a functioning end-to-end self-improving loop with real tool-chain integration. Persisted multiple playbook versions and showed strategy evolution over successive calls. Integrated voice-intelligence signals (not just transcript text) into optimization. Added fallback handling to keep the pipeline running when webhook behavior was inconsistent. Instrumented observability so improvements are inspectable, not just claimed.

What we learned

Voice-level analytics materially improve sales coaching quality over transcript-only analysis. Event-driven systems need idempotency, retries, and visibility from day one. “Autonomous” systems require explicit verification checkpoints between each stage. Strong observability turns a demo into something trustworthy and diagnosable.

What's next for Playbook

Harden reliability (queue-based processing, signed webhook validation, replay-safe events). Add explicit per-call → per-version attribution and automated performance scoring. Expand Lightdash dashboards for conversion, engagement-by-version, and objection-type outcomes. Move from hackathon prototype to production-grade autonomous sales optimization.

Built With

• airia
• elevenlabs
• lightdash
• modulate
• next.js
• supabase