🛡️ Recoup AI

Self-improving, FDCPA-compliant AI voice agents for debt collection — with a Salesforce-style CRM.

▶ Live demo — recoup-ai.vercel.app · Built for the YC Voice Agents Hackathon

📞 Real voice, live: click "Call me" on the site, enter a US/Canada number, and Riley (the live NVIDIA-Nemotron Pipecat agent on Pipecat Cloud) calls your phone via Twilio and runs a fully compliant collections call.

Pipecat · NVIDIA Nemotron · Twilio · Cekura · Next.js 16 · Vercel

🎥 Demo walkthrough — click to play the full 1:40 video

_{▶ Watch the 1:40 walkthrough (plays in GitHub's video viewer) · direct MP4}

---

The one-liner

AI voice agents that call debtors, negotiate a promise-to-pay, and stay 100% compliant — and that continuously self-improve, because every release is graded by a simulated-caller eval suite (Cekura) that feeds failures back into the agent until it hits 100% compliance.

This repo ships two halves:

1. The voice agent (server/) — a Pipecat pipeline (NVIDIA Nemotron ASR + LLM, Magpie/Gradium TTS) that places compliant outbound collections calls over Twilio. Agent "Riley" verifies the right party, delivers the mini-Miranda, negotiates a payment plan, and hard-stops on disputes / cease requests.
2. The CRM (web/) — a Salesforce-Lightning-style operator cockpit to ring outstanding debts, watch FDCPA compliance light up in real time during the call, and prove the agent auto-improves from 62% → 100% compliance via Cekura.

Why collections is the wedge

Collections is manual, expensive, inconsistent, and a legal minefield. A single non-compliant line — a threat, disclosing a debt to the wrong person, calling after 9pm — is a regulatory liability. Teams can't scale coverage without scaling risk, and today they tune scripts on vibes and hope.

Recoup AI doesn't ship "a voice agent that works in a demo." It ships a system that proves it works and gets better on its own.

What's in the CRM

Dashboard — portfolio KPIs (outstanding, promise-to-pay rate, recovery, 100% compliance), aging, an "accounts to call next" queue, and a live activity feed.
Accounts — a sortable, filterable debtor book with status & risk pills and one-click calling.
Account record — record header, action bar (Call now, log PTP, mark disputed, honor cease), tabbed Activity / Details / Script / Compliance, and a right side-panel with full customer context + "what happened" history.
Live call console — dial → ring → connected, a streaming transcript, a real-time FDCPA compliance checklist, a script-step tracker, and automatic outcome capture.
Compliance Lab — Cekura's simulate → evaluate → auto-improve loop: the 62% → 100% climb, 8 adversarial personas, and Claude's per-iteration prompt fixes.

The self-improvement loop (the centerpiece)

                       ┌──────────────────── Pipecat pipeline ────────────────────┐
 Debtor ☎ ⇄ Twilio  ⇄ │ Nemotron ASR → Nemotron-3-Super LLM → Magpie/Gradium TTS │
                       └───────────────────────────┬──────────────────────────────┘
                                                    │  outcome + transcript
                                                    ▼
        ┌──────────────── Cekura auto-improvement harness ───────────────┐
        │ debtor personas → place calls → grade (compliance + recovery)   │
        │      ▼ failures → Claude edits system prompt → redeploy → re-run │
        └────────────────────────  until 100% compliant  ────────────────┘

Cekura runs adversarial debtor personas (wrong party, angry, disputes, "stop calling me", hardship, identity refusal, voicemail…) against the agent, grades every call on compliance + effectiveness metrics, and Claude rewrites the agent's prompt and redeploys until the suite hits 100% compliance and a higher promise-to-pay rate. The Compliance Lab visualizes that climb — 62% → 78% → 94% → 100% across 3 auto-iterations, 0 human edits.

Compliance, enforced (not hoped for)

The agent's hard rules map 1:1 to the FDCPA and are verified continuously by Cekura:

• Right-party contact before any disclosure (§1692b) · Mini-Miranda (§1692e(11))
• Accurate balance (§1692e(2)) · No harassment / threats (§1692d)
• No third-party disclosure (§1692c(b)) · Honor cease-communication (§1692c(c))

🏆 Built for the rubric

The hackathon guidance (preserved in docs/STARTER.md) asks for four things — "Build something new… Use the tools from Cekura to evaluate and improve the performance of what you build. Use Pipecat as the orchestration framework… we encourage you to use the open source models from NVIDIA," and notes the judges "want to see great examples of using Cekura to improve voice agent performance, and using open source models from NVIDIA." Here's how Recoup AI delivers each:

What the judges are looking for	How Recoup AI delivers
Use Cekura to evaluate and improve your agent	The product is a Cekura improvement loop. Riley is graded by 8 adversarial debtor personas on compliance + effectiveness metrics; failures feed back into the system prompt and the agent re-runs until 62% → 100% compliance (PTP rate 41% → 72%), in 0 human edits — visualized live in the Compliance Lab.
Use Pipecat as the orchestration framework	The agent is a Pipecat pipeline end-to-end (STT → LLM → TTS + direct function tools), telephony over Twilio, deployable to Pipecat Cloud (server/pcc-deploy.toml; live as flower-bot.deon).
Use open-source NVIDIA models	The headline build (server/bot-nemotron.py) is 100% NVIDIA — Nemotron Speech Streaming ASR + Nemotron-3-Super-120B LLM + Magpie TTS. A drop-in GPT-4.1 variant (server/bot-gpt.py) shares the exact tools + prompt, so you can A/B NVIDIA vs GPT-4.1 head-to-head inside Cekura.
Build something new — creative, technical, solves a real problem	Collections is a real, painful, regulated problem ($1T+ in US consumer debt). Recoup AI ships a system that proves it's compliant and self-improves — and wraps it in an operator-grade CRM. The reproducible-compliance guarantee is the wedge.

Tech stack

• Voice: NVIDIA Nemotron Speech Streaming (ASR) + Nemotron-3-Super-120B (LLM) + Magpie/Gradium (TTS), orchestrated by Pipecat, over Twilio PSTN, deployable to Pipecat Cloud. Two interchangeable builds — bot-nemotron.py (100% NVIDIA) and bot-gpt.py (GPT-4.1) — share identical tools + prompt for Cekura A/B testing.
• Eval / improve: Cekura simulate → evaluate → auto-improve.
• CRM: Next.js 16 (App Router) · React 19 · Tailwind CSS v4 · TypeScript · framer-motion · lucide-react, deployed on Vercel.

Repo structure

.
├── web/                 # Recoup AI CRM (Next.js, deployed to Vercel) — see web/README.md
├── server/              # Pipecat voice agent — bot-nemotron.py (NVIDIA) + bot-gpt.py (GPT-4.1)
├── PRD.md               # Product spec
├── docs/STARTER.md      # Original hackathon starter instructions
└── docs/screenshots/    # UI screenshots

Run it

CRM (web):

cd web && npm install && npm run dev      # http://localhost:3000

Runs as a self-contained demo with no env vars. See web/README.md to wire real telephony (Twilio → Pipecat) and live Cekura mode.

Voice agent (server):

cd server && uv sync
uv run bot-nemotron.py    # 100% NVIDIA build  (or: uv run bot-gpt.py for the GPT-4.1 build)
# then open http://localhost:7860

Both builds are the same compliant collections agent ("Riley") with identical tools + prompt — run either against Cekura to compare. Full Twilio + Pipecat Cloud + Cekura setup in docs/STARTER.md.

Deploy

The CRM is deployed on Vercel (project root directory web/). The voice agent deploys to Pipecat Cloud (server/pcc-deploy.toml).

---

Built on the YC Voice Agents Hackathon starter (Pipecat + NVIDIA Nemotron). Evaluation & auto-improvement by Cekura.