Watchfloor

Inspiration B2B founders and sales leaders waste countless hours uncovering deal risks too late. Vital warning signals such as a silent buyer, slipping close dates, or a missing economic buyer are buried inside stale CRM fields and activity logs that nobody reviews between meetings. We wanted to build an ambient operations system that turns an Echo Show or personal voice assistant into an intelligent risk desk. Watchfloor was inspired by a simple idea: Deals rarely die in meetings; they die in silence. We wanted to create a system that surfaces these silent risks instantly, explains them with hard evidence, and allows leaders to take action hands-free—with a strict human-in-the-loop guarantee.
What it does Watchfloor is an ambient operations and risk-triage add-on for Alexa+ that monitors B2B sales pipelines: The Morning Briefing: Instantly tells users which deals are slipping and why, rendering a rich triage board on visual displays using MCP Apps. Explainable Risk Engine (RIGS): Computes a deterministic, transparent risk score based on Recency, Intent, Gaps, and Slippage, backing every claim with hard evidence IDs. Asynchronous Deep Triage: Triggers a deep multi-agent review that coordinates specialist agents to analyze timelines, contextualize accounts, and formulate recovery playbooks. Human-in-the-Loop Actions: Proposes actions (like creating tasks or drafting follow-ups) that require explicit user confirmation via secure, single-use tokens before execution. Standing Orders: Persists user intent across sessions, running background checks to push real-time alerts directly into Slack.
How we built it Watchfloor is architected as a modern, resilient, and zero-cost serverless system: Model Context Protocol (MCP): Built a stateless server conforming strictly to the 2025-11-25 Streamable HTTP specification, utilizing FastAPI, Uvicorn, and an AWS Lambda Web Adapter layer with strict DNS rebinding protection (Origin/Host validation). Multi-Tier Inference Fallback: Engineered a custom model provider using the AWS Strands SDK, implementing a dynamic fallback chain that routes from Amazon Bedrock Nova to Google Gemini Flash-Lite upon throttling, ensuring 100% uptime. Asynchronous Orchestration: Deployed Kiro Crew as an autonomous developer workspace and asynchronous engine to handle deep workflows and secure Slack thread sessions. State Management & FinOps: Leveraged Amazon DynamoDB and Postgres for workspace-isolated RLS data storage, utilizing VPC Gateway Endpoints to maintain a 100% Always Free tier cloud footprint.
Challenges we ran into Building an enterprise-grade MCP server within a tight hackathon window came with notable hurdles: Bedrock Quota Provisioning: We encountered the known AWS account provisioning bottleneck where default inference quotas were locked. We overcame this by writing a custom model fallback provider within the Strands SDK to dynamically route through Google AI Studio without breaking our core architecture. MCP Streamable HTTP Compliance: Ensuring stateless compliance, exact error formatting (like -32602 for unknown tools), and proper Server-Sent Events chunking over serverless endpoints required fine-tuning our ASGI middleware and edge guards.
Accomplishments that we're proud of True Protocol Conformance: Built a fully conformant MCP 2025-11-25 server supporting both user and agent scope-filtering, validated against a rigorous test suite. Zero-Cost Production Architecture: Maintained a $0 cloud footprint by optimizing Always Free AWS services and implementing smart FinOps strategies (avoiding costly NAT gateways). Robust Multi-Agent Resilience: Successfully decoupled our triage agents from blocked cloud runtimes, establishing a multi-tier fallback framework that ensures the demo never goes dark.
What we learned Working with the Model Context Protocol and cutting-edge agentic toolkits taught us the immense value of decoupling language understanding (handled by Alexa+) from deterministic execution (handled by our high-speed tool backend). We also gained deep practical insights into writing resilient error-handling logic for serverless AI applications.