EthiCo: The AI Compliance Copilot for Debt Collection


Inspiration

Debt collection calls are high-stakes.
A rushed phrase or missed disclosure can breach FDCPA or Reg F, triggering both regulatory risk and customer distress.
We wanted something better than post-call audits — a live AI partner that helps agents stay compliant while the call is still happening.
That’s the spark behind EthiCo.


What It Does

EthiCo is an AI compliance copilot that sits alongside every agent:

  • Streams the live call, transcribes it in real time, and mirrors the transcript in the dashboard.
  • Each new utterance is routed through EthiCo’s backend, which checks for risky language and flags potential violations instantly.
  • Sensitive moments trigger on-screen warnings with context, timestamps, and the specific rule that might be at risk.
  • Supervisors and agents can access both the Live Chat View (transcript + alerts) and the Agent Analytics View to review calls after they end.

How We Built It

  • Frontend (Vite + React)
    Two dashboards:

    • Live transcript with violation badges
    • Analytics pane with warnings/history
      Includes an upload & simulation mode for recorded calls.
  • Backend (Flask + Socket.IO)
    Manages session lifecycle, streams audio chunks, finalizes transcripts, and stores call logs in MongoDB.

  • Speech-to-Text
    ElevenLabs STT for accurate diarized transcripts (with simulated-mode fallback for demos).

  • LLM Layer
    Google Gemini API classifies each utterance (live) and whole transcripts (upload/simulation) as
    violation / sentiment / OK, supported by cached knowledge snippets.

  • Knowledge Base
    kb_chunks collection in MongoDB seeded via SerpAPI + Trafilatura.
    Embeddings generated with BGE-M3 and reused for retrieval.

  • Storage
    All agents, customers, sessions, and call records live in MongoDB Atlas,
    so uploads, live calls, and simulations share the same audit trail.


Challenges We Ran Into

  • Getting reliable transcriptions fast enough for live monitoring — fallback to simulation/upload ensured full coverage when mic access or STT quotas failed.
  • Maintaining consistent session IDs between live, simulation, and upload flows — refactored to make every path produce a traceable call record.
  • Aligning frontend timing with backend classifications (live warnings vs. whole-call analysis) to ensure transcripts never “jump” or lose warnings mid-session.

What We Learned

  • Real-time compliance isn’t just LLM prompts — it’s session orchestration (STT buffers, socket events, DB writes) so every warning has a source and timestamp.
  • Knowledge summaries + cached embeddings keep latency low without sacrificing legal accuracy.
  • UX matters — agents need alerts that are clear, cite the rule, and fit seamlessly alongside the transcript they’re reading.

What’s Next for EthiCo

  • Prosody cues (tone/volume spikes) to complement text-based alerts.
  • State/regional rule packs to toggle FDCPA + state-specific (e.g., Rosenthal, NY DFS) guidance per call.
  • Supervisor dashboards showing aggregate risk trends, top violations, and coaching opportunities.
  • Enterprise connectors (CRM, dialers) so EthiCo’s session IDs and transcripts sync automatically into existing workflows.
  • Auditability: exportable logs + human verification loops for every warning so compliance teams can review or override.

Closing Line for Judges

“EthiCo turns compliance into a live companion — catching risky language at the exact moment it’s spoken, preserving the call record, and giving supervisors proof that every conversation stayed on the right side of the law.”

Built With

Share this project:

Updates