EKAM landing in dark mode with responsive audio ring and real-time MCP capability highlights.
Setting up family member roles and greeting preferences for tailored living-room responses.
Kids homework tracking alongside an immutable, transparent safety audit ledger for every tool call.
Quick family onboarding to personalize preferences, memory retention, and parental controls.
EKAM landing experience in light mode: One natural sentence orchestrates the entire family evening.
Persistent conversational memory: EKAM recalls past interactions so you never have to repeat yourself.
Grandma check-in reminder card and family calendar conflict schedule powered by MCP tools.
Multi-language support (EN/Hindi/Japanese) with a safe two-phase commerce proposal gate.

Inspiration

Modern family evenings are chaotic and fragmented across isolated apps: parents check schedules, kids demand TV shows, delivery couriers ring the doorbell, snacks need to be ordered, and elderly grandparents need caretaking check-ins. We wanted to build an agentic experience where a single, natural sentence—"movie night with kids"—coordinates the entire evening across smart home devices, media, commerce, and family caretaking without touching a phone.

What it does

EKAM orchestrates household evening routines through a multi-tool agentic loop backed by an open Model Context Protocol (MCP) server. Saying "movie night with kids" triggers 6 coordinated actions simultaneously:

Calendar Check (get_calendar)**: Verifies evening family schedules and homework deadlines.
TV Control (tv_control): Turns on the living-room TV, launches family streaming, and locks **Kids Mode ON to family-safe titles (Bluey, Encanto).
Front-Door Camera (door_status): Checks the smart porch camera for visitors or package deliveries.
Living-Room Scene (scene_control): Dims ambient lighting to 20% for movie night.
Safe Commerce Proposal (propose_order): Stages popcorn, samosas, and juice with a parental PIN confirmation gate—no money moves without explicit voice approval.
Elderly Care Check-In (remind_grandma): Automatically schedules a morning video call with Grandma.

EKAM renders living-room-ready visual cards (Media Player, TV Mockup, Camera Viewport with scanlines, and Order Tracker) and speaks responses aloud using Amazon Polly Neural TTS.

How we built it

Agent Core & Reasoning (agent/agent.py): Built with a Strands-style agent loop. Integrated with Amazon Bedrock via boto3 bedrock-runtime supporting both Amazon Nova Micro and Anthropic Claude formats with multi-turn memory context.
MCP Server (mcp-server/server.py): A pure Python JSON-RPC 2.0 server adhering to the official MCP 2025-11-25 Streamable HTTP specification, exposing 11 smart home and commerce tools.
Voice & Speech (api/main.py): Powered by Amazon Polly neural text-to-speech engine (POST /api/speak) with voices like Joanna and Aditi.
Memory & Audit Trail: Multi-turn family dialog memory (agent/memory.json) and an append-only safety audit ledger (data/audit.jsonl).
Simulated TV UI (app/): High-performance, zero-build vanilla HTML5, CSS3, and JavaScript served via FastAPI.

Challenges we ran into

Graceful Fallback & Zero-Fail Guarantee: When running in sandbox hackathon environments without active AWS credentials or Bedrock model quota access, the agent needed to degrade seamlessly to deterministic local reasoning so judges never experience a crash.
Safe Purchasing Guardrails: Preventing accidental orders by children required separating tool execution into a two-phase commit: propose_order stages the cart, and confirm_order requires a parent voice PIN.

Accomplishments that we're proud of

11 Real MCP Tools: Clean separation of protocol and business logic following the official MCP 2025-11-25 specification.
Zero Build Step: Pure Python 3.9 + vanilla web standards that launch instantly anywhere with zero npm or webpack dependencies.
Family-Centric Safety: Multi-turn persistent memory, kid-safe shields, and elderly check-ins running entirely on a serverless architecture.

What we learned

Building with Model Context Protocol (MCP) demonstrated how standardized tool interfaces make AI agent tool dispatching modular and future-proof. Integrating Amazon Bedrock showed how lightweight models like Nova Micro can polish conversational replies while keeping latency minimal.

What's next for EKAM

Integrating live IoT hardware protocols (Matter / Zigbee).
Voice biometrics to automatically distinguish parental approval from children's requests.
Real-time video stream analysis from front-door cameras via AWS Rekognition.