IBM_Hackathon_Report/problem_statement.md

Problem & Solution Statement

Problem

In industries where critical decisions depend on the accuracy of verbal communication—such as finance, healthcare, legal compliance, and customer support—important claims made during meetings or calls often go unverified.
For example:

• A customer service rep claims a service has "99.99% uptime."
• A vendor promises "P95 latency under 200 ms globally."
• A team assures compliance with "30-day data retention policies."

These statements, if untrue or unsupported, can result in:

• Legal liabilities
• Financial loss
• Reputation damage
• Regulatory penalties

Existing meeting transcription and summarization tools do not go far enough — they focus on creating readable notes, but do not actively validate or cross-check claims against authoritative knowledge bases.

---

Solution

We have built ClaimCheck AI, an agentic AI system that automatically:

1. Transcribes conversations (speech-to-text).
2. Extracts claims made in the dialogue.
3. Retrieves relevant evidence from a curated Knowledge Base (KB) using IBM watsonx embeddings and reranking.
4. Verifies whether each claim is supported, contradicted, or unverifiable.
5. Summarizes the findings in a structured, auditable report.

---

Target Users

• Compliance Officers – Automatically flag unsupported statements in regulated industries.
• Enterprise Managers – Validate vendor promises and internal KPIs.
• Customer Success & Support Teams – Ensure commitments to customers are accurate.
• Legal Teams – Use as supporting evidence for contractual disputes.

---

User Interaction Flow

1. Upload a call recording or connect via Zoom/phone integration.
2. The platform transcribes the audio.
3. An AI claim extractor detects possible factual statements.
4. IBM watsonx-powered retrieval finds the top-matching KB evidence.
5. A verifier agent cross-checks claims against evidence.
6. A final report is generated with:
   • List of claims
   • Associated evidence
   • Verdict: Supported / Contradicted / Unverifiable

---

Creativity & Uniqueness

Unlike traditional meeting summarizers:

• Evidence-backed summaries — not just "what was said," but "whether it’s true."
• Multi-agent pipeline — dedicated agents for transcription, claim extraction, retrieval, verification, and summarization.
• Scalable KB integration — supports industry-specific compliance rules, SLAs, and historical records.
• Audit trail — every verdict links back to the evidence snippet.

This approach moves beyond passive transcription to active truth verification in real time.

---

Agentic AI in Action

ClaimCheck AI works as a multi-agent system:

1. Transcriber Agent – Converts speech to text (IBM Speech-to-Text possible).
2. Claim Extractor Agent – Uses LLM prompting to detect factual statements.
3. Retriever Agent – Uses IBM watsonx text embeddings + FAISS + IBM reranker.
4. Verifier Agent – Cross-references claims with evidence and generates verdicts.
5. Summarizer Agent – Produces structured human-readable reports.

---

High Impact

ClaimCheck AI could:

• Prevent false compliance reporting in regulated industries.
• Detect SLA breaches before they become customer escalations.
• Save legal teams countless hours in discovery and evidence-gathering.