You said it out loud. You didn't realize it was a signal.
Every VC tracks metrics. Nobody captures what they hear.
Second Listen is a voice agent for post-investment risk management. After a founder call, board meeting, or site visit, the investor talks through what happened at a natural pace. The agent, powered by a structured risk playbook, listens for risk signals hidden in small talk, asks checklist-driven follow-up questions, and turns the debrief into evidence tables, ledger updates, and escalation flags.
Built for the lablab.ai × AssemblyAI Voice Agent Hackathon (Sep 1–30, 2026).
Post-investment risk signals come in two channels:
- Dashboard signals (revenue decline, runway < 6 months) — visible in KPI reports, already covered by portfolio monitoring tools.
- Conversation signals (key-person departure, funds used outside the agreed purpose, litigation) — they only surface in conversations. By the time quarterly review forms get filled, memory has eaten half the evidence.
Second Listen covers the second channel: the gap between "the conversation happened" and "the signal is in the ledger".
- Debrief — after a call, press and talk for as long as the account needs.
- Interrogate — the agent follows the playbook's escalation checklist ("Who's covering finance since the CFO left? Was there written approval for moving the grant money?"), one question at a time.
- Record — produces an evidence table with session-relative capture times, a risk-ledger update, escalation flags per policy, and a downloadable Markdown follow-up note per company. The AI surfaces evidence and suggestions; authorized people make the final risk decision.
Prefer not to talk? Analyze recording uploads a pre-recorded debrief (Chinese works too): Universal-2 transcribes it, then the same evidence framework returns candidate signals and follow-up questions. With an OpenAI-compatible LLM configured it performs the full upload review; otherwise a deterministic keyword fallback marks candidates for human review.
The 3:44 demo uses real product runs: a second check-in call that reopens last week's follow-ups one at a time, live evidence capture with the investor's own quotes, and an upload-mode analysis of a recording. One quiet section of the live call was removed for length; the live segments otherwise run at their recorded pace. Narration script: docs/hackathon-demo-script-2min.md.
- AssemblyAI Voice Agent API — one WebSocket for STT + LLM + TTS, turn detection, barge-in, tool calling
- Playbook skill: a structured risk framework (evidence dimensions, escalation checklist, and human confirmation) injected as the agent's system prompt
- Backend is Python standard library only —
requirements.txtinstalls nothing; audio transport runs in the browser (AudioContext+AudioWorklet) directly against the AssemblyAI WebSocket
This repository includes a root-level render.yaml. In Render, create a Blueprint from this repository and provide ASSEMBLYAI_API_KEY. The service runs the real browser application over HTTPS; its server-only API key mints single-use 60-second tokens, and every live session is capped at five minutes. The public contest demo also permits at most two token requests per visitor and six total per 15 minutes; recording uploads are limited to one per visitor and three total per hour. Optional LLM_BASE_URL, LLM_API_KEY, and LLM_MODEL variables enable the full upload-review path.
The free Render plan is suitable for a short trial, but it sleeps when idle and its filesystem is ephemeral. Use a paid persistent disk mounted at /var/data, or an external data store, if ledger and cross-debrief history must survive restarts.
Windows: double-click 启动SecondListen.bat in the repo root, then open http://localhost:3000.
Any OS with Python 3.9+:
cd app
cp .env.example .env # add your ASSEMBLYAI_API_KEY
AGENT=second-listen python publish.py
python deployment/browser/server.pyDeployment and private-trial options are in TEAM.md.
Each company keeps one small JSON history file (app/data/history/<company>.json). The next debrief for that company opens by checking the previous session's follow-ups — owner, deadline and all — one at a time, and skips re-asking signals already recorded. No database, no vector store; details in docs/roadmap.md.
Second Listen's project-authored code and documentation are released under the MIT License. The playbook skill keeps its original MIT notice in skill/LICENSE.
The app/ runtime began from AssemblyAI's official Voice Agent starter repositories and has since been adapted for Second Listen. Upstream provenance and the project-specific contribution boundary are recorded in NOTICE; third-party portions remain subject to their owners' terms.