Parallax

"PARALLAX is a standout submission that demonstrates genuine Band SDK integration for a high-stakes, real-world use case: cyber-fraud investigation for banking malware. This is a GenAI-native platform that analyzes suspicious Android applications to identify banking-malware behavior (accessibility abuse, overlay attacks, SMS interception) and converts findings into structured evidence, confidence-scored claims, and investigation artifacts for banks, fraud teams, and SOC analysts. The Band SDK integration is **legitimate and well-executed**. Eight specialist agents (Intake, Device Compromise, Transaction Trace, Mule Graph, Evidence Validator, Liability, Legal Evidence, Decision Convenor) run inside `band.Agent` processes and connect to Band rooms via the Python SDK. The README explicitly documents the integration: "PARALLAX posts an immutable evidence bundle... into the room and @mentions each agent. They post claims, the Evidence Validator challenges weak ones, the Decision Convenor refuses to mark the packet `final` until every challenge resolves." This is proper @mention-driven coordination with agents as Band room participants. The team transparently acknowledges using Band's Pro plan agent-side endpoints (`/api/v1/agent`) rather than Enterprise chatrooms API, with a fallback to local transcript export for Human API requirements. The feat/band-integration branch contains the implementation, and the demo video shows the Band chat interface with @PARALLAX agents collaborating in real-time. The domain choice is excellent: cyber-fraud investigation is a critical, high-stakes enterprise workflow where multi-agent evidence validation and human-in-the-loop decision-making are essential. The eight-agent specialization mirrors real SOC/fraud investigation workflows: intake triages reports, multiple agents analyze different malware indicators (device compromise, transaction traces, mule accounts), Evidence Validator challenges weak claims, and Decision Convenor synthesizes the final packet only after all challenges resolve. This is sophisticated multi-agent orchestration with adversarial validation built in. The technology stack is production-grade: LangChain, CrewAI, LlamaIndex, Qdrant (vector DB), Redis, OpenAI GPT-4o, Mistral, Phi-3, LLavA, Claude. The system performs static/dynamic analysis of APKs, extracts IOCs, assesses permissions, and produces exportable evidence bundles. The "live investigation modeling" approach moves beyond static reports to provide auditable reasoning trails—critical for regulatory compliance and legal proceedings in fraud cases. Presentation is strong: the demo video shows the Band interface with agent @mentions and collaboration, clear evidence of the investigation workflow, and professional production values. The GitHub repo has 23 deployments (production + preview), showing operational maturity. Business value is exceptionally high: banking fraud costs billions annually, and automating evidence-gathering while maintaining audit trails addresses a genuine enterprise pain point. Originality is strong: while malware analysis tools exist, the specific combination of Band multi-agent orchestration with adversarial evidence validation (Evidence Validator challenges claims), human-controlled final decisions, and structured legal evidence generation for fraud cases is novel and shows creative problem-solving. This is one of the strongest Band SDK implementations in the hackathon, demonstrating proper understanding of agent registration, @mention coordination, and leveraging Band's audit trail for regulated workflows."