Cascade

• 

  Real-time multi-modal route tracking and active corporate passenger monitoring.

• 

  Automated travel preferences, risk tolerance, and strict corporate policy guardrails.

• 

  Orchestrating first-mile, main hub-to-hub transit, and destination ground transfers.

• 

  Real-time event injection and autonomous multi-agent reasoning logs.

• 

  Verified resolution metrics, SLA latency tracking, and cryptographic transaction logging.

• 

  Side-by-side view of disrupted itinerary vs. autonomously rebooked route.

• 

  Instant mobile passenger alerting and automated disruption resolution dispatch.

Inspiration

When an executive experiences a 45-minute flight delay, the downstream domino effect is instantaneous and catastrophic: missed high-speed train connections, invalidated hotel check-in windows, and stranded personnel missing critical business commitments.

Legacy travel platforms leave passengers trapped in multi-hour customer service queues because flights, rail, and hotels operate in disconnected data silos. We built CASCADE to eliminate this friction entirely: an autonomous, self-healing travel recovery engine that detects disruptions instantaneously via Change Data Capture (CDC), reasons across multi-modal alternatives with generative AI, and commits resilient rebookings with distributed ACID guarantees in sub-second time.

What it does

CASCADE is an autonomous executive travel and multi-modal logistics recovery engine powered by CockroachDB Serverless and AWS Bedrock:

🧠 Autonomous Multi-Agent Reasoning

• Instant disruption detection: Ingests live delay and cancellation events via CockroachDB Change Data Capture (CDC) and triggers an autonomous 10-step multi-agent recovery workflow.
• 1536-Dimensional vector preference recall: Utilizes Amazon Titan Embeddings stored in CockroachDB’s native VECTOR(1536) columns with HNSW cosine-similarity indexing to recall traveler preferences (cabin tiers, layover thresholds, rail vs. air affinity).
• Multi-Branch candidate evaluation: Simultaneously evaluates competing paths:
  • Branch alpha: Speed priority (fastest transit time).
  • Branch beta: Cost optimization (carrier-covered / zero cost delta).
  • Branch gamma: Preference alignment (personalized HNSW vector winner).
• Smart feeder bypass: Identifies feeder delays (>45 min) and reroutes passengers to High-Speed Rail (HSR) or executive transfers to safeguard critical international legs.

🛡️ Enterprise governance & reliability

• Human-in-the-Loop (HITL) guardrails: Autonomously executes low-friction rebookings within policy thresholds ($300 limit), while instantly escalating policy deviations to managers for one-click approval.
• Distributed ACID & SAGA rollbacks: Commits bookings under strict SERIALIZABLE transaction isolation in CockroachDB with automatic 40001 retry handling and SAGA compensation for multi-segment itineraries.
• Immutable Audit Trails: Generates SHA-256 cryptographic proofs and downloadable compliance PDF reports for every rebooking decision.

📲 Real-Time experience & telemetry

• Live stream dashboard: Streams agent reasoning steps and fleet topology graphs via Server-Sent Events (SSE).
• Telegram mobile companion (@CascadeAWS_bot): Dispatches rich interactive disruption alerts, rebooking cards, and instant transaction hashes directly to the traveler's phone.

How we built it

• Distributed database & vectors: CockroachDB Serverless - native VECTOR(1536) storage, HNSW indexing, foreign key topologies, CDC changefeeds, and SERIALIZABLE transaction isolation.
• AI orchestration & reasoning: AWS Bedrock (Anthropic Claude 3.5 Sonnet & Amazon Titan Embeddings v2) coordinated via the Model Context Protocol (MCP) TypeScript SDK with 6 specialized database and routing tools.
• Backend infrastructure: Node.js & TypeScript 5.4, Express, Server-Sent Events (text/event-stream), Zod schema validation, and PDFKit for automated audit exports.
• Frontend & visuals: Vanilla HTML5 & Glassmorphism UI featuring a 3-step Journey Builder, live fleet topology visualizers, and interactive Chaos Simulators (Contention Engine, Multi-Region Failover, Triple-Failure Mode).
• Mobile gateway: Telegram Bot API (telegraf) for real-time traveler notifications and deep-link approvals.

Challenges we ran into

1. Distributed serialization contention (40001): High-contention scenarios where multiple travelers compete for the last seat caused concurrency conflicts. We built an exponential backoff-and-retry wrapper around CockroachDB's SERIALIZABLE transactions with deterministic priority weighting.
2. Sub-Second agent orchestration: Multi-agent LLM reasoning can introduce latency. We kept vector similarity search directly inside CockroachDB via native HNSW indexes eliminating external vector DB network hops and streamed step telemetry over SSE.
3. Multi-Modal graph dependencies: Travel segments (flight ➔ train ➔ hotel ➔ transfer) have strict sequential dependencies. We designed a transactional linked graph schema and SAGA compensation coordinator to guarantee clean rollbacks if an intermediate leg failed.
4. Bidirectional MCP Tool-Calling: Engineering robust, type-safe communication between the Model Context Protocol (MCP) server, AWS Bedrock runtime, and live CockroachDB execution tools.

Accomplishments that we're proud of

• Sub-Second SLA: Autonomous execution from CDC event trigger to AI evaluation, database commit, cryptographic proof generation, and Telegram notification in under 1 second.
• Unified data tier: Eliminated external vector databases and caching layers by relying entirely on CockroachDB for relational data, ACID guarantees, and 1536-dimensional HNSW search.
• Production-Grade resilience: Built automated SAGA rollbacks, HITL corporate guardrails, and cryptographic PDF audit proofs into the core transaction loop.
• Graceful offline fallback: Designed deterministic heuristic fallbacks ensuring the system remains operational and demonstrable under any network or credential constraint.

What we learned

• The operational simplicity and performance of collocating relational data, Change Data Capture, and native vector search (VECTOR + HNSW) inside CockroachDB.
• Designing resilient, structured tool pipelines using the Model Context Protocol (MCP) and Claude 3.5 Sonnet on AWS Bedrock.
• Implementing SAGA compensation patterns to prevent partial state corruption across multi-leg transportation networks.

What's next for Cascade

• Commercial pilot launch: Transitioning from our hackathon deployment to real-world corporate pilots, integrating directly with enterprise ERPs (SAP Concur, TravelPerk) and GDS networks (Sabre, Amadeus).
• Next-Gen mobile & voice experience: Expanding the Telegram companion into a dedicated cross-platform mobile application featuring real-time AI voice assistance and one-tap boarding pass provisioning via Apple Wallet.
• Predictive radar rebooking: Integrating live FAA/Eurocontrol air traffic telemetry and Doppler weather data to initiate proactive rebookings before airlines officially declare cancellations.
• Multi-Region Geo-Partitioning: Scaling CockroachDB across us-east-1, eu-west-1, and ap-southeast-1 with localized data residency and sub-millisecond edge latency worldwide.

Built With

• api
• aws-bedrock
• cockroachdb
• css
• express.js
• github
• heroku
• html
• javascript
• llm
• multi-agent
• node.js
• typescript