RecoverX — AI Revenue Recovery Agent

Part of Harshil Desai's evidence-first tools for Indian finance: LoanLens · PaisaTrail · IPO Lab · RecoverX

RecoverX is a safety-first revenue recovery MVP built for the Razorpay AI Revenue Recovery track. It uses an LLM to interpret customer context, then routes that context through deterministic policy guardrails before any recovery action is selected.

Important: payment outcomes and recovered amounts in this repository are simulated for the MVP benchmark. RecoverX does not move real money.

Why RecoverX

Traditional recovery systems often retry failed payments without considering what the customer is saying. RecoverX adds customer context while keeping financial action deterministic and auditable.

Customer reply
      ↓
AI / safety-first intent interpretation
      ↓
Deterministic policy guardrails
      ↓
RETRY · PAYMENT LINK · REMINDER · FOLLOW-UP · ESCALATE · STOP
      ↓
Deterministic simulated outcome
      ↓
Audit trail

The LLM does not execute financial actions directly.

Core guardrails

Signal	Policy action
Dispute / already-paid claim	STOP
Fraud / unauthorized-use claim	ESCALATE
Retry limit reached	STOP
High-value transaction	ESCALATE
Promise to pay later	FOLLOW-UP
Temporary network error	RETRY
Expired card	PAYMENT LINK

Current benchmark

The bundled synthetic dataset contains 47 transactions with ₹6,52,305 total value at risk.

A representative controlled run produced:

Metric	AI-assisted	Rules-only
Recovered	₹1,16,363	₹1,16,363
Automated recovery attempted	₹2,38,316	₹2,72,267
Attempt success rate	48.8%	42.7%
Follow-up value	₹59,558	₹0

The benchmark therefore demonstrates recovery parity with lower automated exposure and better successful-attempt efficiency, rather than claiming unsupported additional recovered revenue.

Product highlights

• Customer-context intent classification with persistent caching for reproducible demos.
• Deterministic policy engine that keeps hard financial guardrails authoritative.
• Deterministic paired outcome simulation so the same transaction gets the same underlying outcome score across AI and rules runs.
• Recovery queue for promise-to-pay cases.
• Transaction-level decision explanation and audit trail.
• CSV audit export.
• Automated safety tests.

Architecture

Run locally

python -m venv venv
venv\Scripts\activate
pip install -r requirements.txt

Create .env in the project root:

GROQ_API_KEY=your_api_key_here

Then:

python app.py              # debug off by default
$env:FLASK_DEBUG=1; python app.py   # opt-in debug for local demos

Open http://127.0.0.1:5000.

Tests

python -m unittest discover -s tests -v

The safety suite (run on every push via GitHub Actions) covers dispute hard-stop, fraud escalation, promise-to-pay follow-up, retry limits, guardrail precedence over AI intent, aged-receivable escalation, follow-up accounting, and deterministic simulation.

Demo flow

1. Run with AI ON.
2. Show the impact panel and follow-up queue.
3. Open a promise-to-pay case and show intent → policy → follow-up.
4. Open a dispute case and show the hard STOP guardrail.
5. Toggle AI OFF and compare the same batch.
6. Export the audit trail.

Tech stack

• Python 3.11+
• Flask
• Groq API (openai/gpt-oss-20b)
• Vanilla JavaScript + Chart.js
• CSV / JSON / JSONL audit data

Security notes

• Keep .env out of Git.
• Runtime cache and audit-history files are ignored by Git.
• Use a production WSGI server and disable Flask debug mode outside local demos.

License

MIT