Dispute Desk
Autonomous multimodal chargeback representment agent with zero-hallucination citation gating.
"Turns chargeback disputes into evidence-backed responses, with every claim traced to a source."
What it is
Built solo for the Razorpay AI Buildathon 2026 (Track 02: AI Risk Manager). Dispute Desk is an autonomous agent that ingests complex card-network chargebacks (Visa, Mastercard, RuPay), extracts transactional evidence across auth logs and customer communication history, and drafts audit-proof representment packages that operations teams ordinarily write by hand.
Instead of generating text at the tail end of an unconstrained chat completion, Dispute Desk places the LLM directly inside the decision loop, governed by strict Pydantic JSON schemas and a zero-hallucination citation grounding gate.
System Architecture
- Multimodal Ingestion Engine: Parses incoming dispute packets—including scanned PDF chargeback notices, transaction timelines, card-network reason codes, and dispute deadlines.
- AI Studio Reasoning & Prompt Caching: Employs Gemini 1.5 with cached system instructions detailing card-network dispute regulations (e.g. Visa Compelling Evidence 3.0 rules).
- Structured Function Calling: Maps dispute requirements to deterministic tool executions (e.g.
query_auth_trace(),lookup_rule_requirements(),validate_avs_cvv_match()). - Two-Stage Grounding Gate: Every claim or citation in the final representment letter is cross-checked against the ingested evidence corpus. Any ungrounded assertion is flagged and rejected before submission.
What was hard
Card-network dispute representments require absolute precision: citing the wrong transaction identifier, hallucinating a shipping tracking number, or invoking an obsolete reason code results in immediate financial loss and scheme fines. Enforcing deterministic grounding without degrading model fluency required implementing a two-stage verification pass: first generating structured evidentiary claims, then running an automated citation-checker that matches every factual token back to verified raw database payloads.
What was measured
Tested across 100 benchmark dispute packets spanning card-absent fraud, merchandise not received, and duplicate processing claims:
- Zero Fabricated Citations: 100% of generated evidence items were strictly grounded in telemetry logs.
- Sub-5-Second Generation: Mean pipeline turnaround of under 4.2 seconds from packet ingestion to ready-to-file representment PDF.
- Shipped with reproducible evaluation scripts, unit tests, and a 5-minute end-to-end video walkthrough.
Empirical Claims & Verification Invariants
| statement | value | sample | status | provenance |
|---|---|---|---|---|
| Grounding gate cross-verifies all citations against source payment telemetry, achieving zero fabricated claims. | 0 fabricated citations | 100 test dispute packets | Measured | eval/grounding_verification_report.json |
| End-to-end multimodal chargeback ingestion, timeline synthesis, and formal response drafting latency. | < 4.2 s | Full representment dossier | Measured | benchmarks/latency_profile.json |