GENIUS — Graph Engine for Networked Infrastructure Understanding & Service

The Home World Model for the Alexa+ Era
Submitted to the Amazon Developer Hackathon 2026 (Primary Track: Alexa+ | Mini-Challenges: AWS Builder & Open Source)

Live AWS Simulator · beepdb Open Source Library · Devpost Submission

---

Executive Summary: What Is Real vs. Simulated

GENIUS (Graph Engine for Networked Infrastructure Understanding & Service) is a self-hosted Model Context Protocol (MCP) server (spec 2025-11-25, Streamable HTTP transport at POST /mcp) paired with a browser-based Alexa+ Preview Device Simulator (client/app.js).

What Runs Live in Code

1. Spec-Compliant MCP Streamable HTTP Client & Server: client/app.js connects directly to POST /mcp using JSON-RPC 2.0 (initialize $\rightarrow$ Mcp-Session-Id header $\rightarrow$ notifications/initialized $\rightarrow$ tools/list $\rightarrow$ tools/call) and renders UI cards dynamically from the MCP server's JSON responses.
2. 12 Registered MCP Tools: Voice-native chirp diagnosis (diagnose_from_description), companion-sensor DSP feature extraction (hear_sound + diagnose), persistent SQLite Home Graph (list_entities, get_home_health, register_entity, record_incident), strict propose_action $\rightarrow$ confirm_action gate, warranty ledger (check_warranty, file_warranty_claim), and live Amazon Bedrock Nova Pro (amazon.nova-pro-v1:0) 5-year TCO deliberation (deliberate_repair).
3. Deterministic Signal Processing & Physics: Hilbert transform amplitude envelope, Hann-windowed FFT, two-parameter Weibull hazard reliability ($R(t)$), and lumped-capacitance building heat-loss simulation ($dT/dt$) with 5,000-run Monte Carlo uncertainty quantification.

What Is Simulated (Per Alexa+ Track Guidelines)

• Because the Alexa+ Developer Preview is gated and does not stream raw room audio to third-party MCP servers, the primary Alexa+ interaction path is conversational (diagnose_from_description), while hear_sound represents a companion sensor intake (e.g., mobile app or auxiliary edge microphone).
• Package delivery (POST /simulate/delivery), weather drops (POST /simulate/freeze), and order fulfillment references (AMZN-2026-94819-...) are simulated state triggers used to demonstrate multi-day lifecycles inside a 2-minute evaluation window.
• When the static S3 simulator URL is opened without a local MCP server running, it explicitly displays an amber Offline Replay (No MCP Server Connected) badge.

---

1. Architecture & The 12 MCP Tools

                      +----------------------------------+
                      |   ALEXA+ PREVIEW SIMULATOR       |
                      |   (Browser MCP Client app.js)    |
                      +----------------+-----------------+
                                       |
                                       | 1. MCP Streamable HTTP (2025-11-25)
                                       |    POST /mcp (JSON-RPC 2.0 + Mcp-Session-Id)
                                       | 2. SSE Proactive Push (/events)
                                       v
+--------------------------------------+---------------------------------------+
|                              GENIUS MCP SERVER                               |
|                                                                              |
|   +---------------------+   +---------------------+   +------------------+   |
|   | 12 MCP Tools        |   | Acoustic Pipeline   |   | SQLite Graph     |   |
|   | - hear_sound        |   | - Hilbert Envelope  |   | - Entities       |   |
|   | - diagnose          |   | - FFT Peak (3368Hz) |   | - Incidents      |   |
|   | - diagnose_from_... |   | - beepdb Matcher    |   | - Warranties     |   |
|   | - list_entities     |   | - data.fill(0.0)    |   | - Proposals      |   |
|   | - get_home_health   |   +---------------------+   | - Provenance     |   |
|   | - propose_action    |                             +------------------+   |
|   | - confirm_action    |   +---------------------+                          |
|   | - record_incident   |   | Amazon Bedrock      |   +------------------+   |
|   | - register_entity   |   | Nova Pro Engine     |   | Amazon Polly     |   |
|   | - check_warranty    |   | - TCO Deliberation  |   | Neural TTS       |   |
|   | - file_warranty_... |   | - Tag OCR Vision    |   | /api/polly/speak |   |
|   | - deliberate_repair |   +---------------------+   +------------------+   |
|   +---------------------+                                                    |
+------------------------------------------------------------------------------+

#	Tool Name	Key Parameters	Function & Output
1	hear_sound	audio_clip_b64, location, interval_s	Companion sensor intake. Extracts Hilbert/FFT features in volatile RAM, zeroes buffer (data.fill(0.0)), returns features.
2	diagnose	interval_s, peak_freq_hz, duration_ms	Matches features against beepdb, logs incident with provenance, stages purchase/remedy proposal.
3	diagnose_from_description	interval_s, chirps_per_burst, brand, device_hint	Voice-native Alexa+ path: Matches spoken chirp cadence, burst count, and brand hints against beepdb without raw audio.
4	list_entities	lifecycle_filter	Queries the Home Graph hardware registry with verified provenance tags and Weibull reliability scores.
5	get_home_health	none	Computes composite 0–100 health score, healthy/aging/EOL counts, and MCP Apps card payload.
6	propose_action	kind, title, entity_id	Strict architectural gate: Stages structured proposal in SQLite. Never mutates external state.
7	confirm_action	proposal_id, confirmed	Authorization mutation gate. Marks proposal confirmed and returns dynamic order reference only after explicit user approval.
8	record_incident	entity_id, description, severity	Appends anomaly record to the persistent SQLite Home Graph ledger.
9	register_entity	brand, model, location, year	Day 0 onboarding. Registers equipment specifications (or parses nameplate image via Bedrock Nova Pro) and seeds warranty policies.
10	check_warranty	entity_id	Inspects policy coverage dates, active/expired status, and manufacturer replacement terms.
11	file_warranty_claim	entity_id, reason	Assembles manufacturer replacement claim dossier for devices that fail within warranty.
12	deliberate_repair	entity_id, issue_description	AWS Builder: Invokes Amazon Bedrock Nova Pro (amazon.nova-pro-v1:0) for 5-year repair-vs-replace TCO analysis.

---

2. Signal Processing, Reliability Priors & Thermal Physics

A. Acoustic Feature Extraction (audio/features.py & audio/ephemeral.py)

1. Analytic Signal Envelope (Hilbert Transform): Computes $A(t) = |s(t) + i \cdot \mathcal{H}{s(t)}|$ to measure continuous pulse width ($\text{ms}$) across oscillating $3\text{ kHz}+$ piezo square waves without zero-crossing dropouts.
2. Spectral Peak Extraction (Hann-Windowed FFT): Extracts dominant frequency $f_{\text{peak}}$ above $400\text{ Hz}$. Physical smoke detector piezo recordings (Bloofrzo CC0 #819808 at $3,368.1\text{ Hz}$ and #819807 at $3,122.3\text{ Hz}$ vs $3,200\text{ Hz}$ nominal manual spec) motivated our calibrated $\pm 350\text{ Hz}$ tolerance window in beepdb.
3. Zero Disk Persistence: Raw audio is never written to disk, logs, or SQLite (verified by test_audio_never_persisted_to_disk_or_db), and in-memory PCM arrays are cleared best-effort via data.fill(0.0).

B. Continuous Reliability Priors (graph/store.py)

Equipment lifecycle degradation is modeled using two-parameter Weibull priors:

$$R(t) = \exp\left(-\left(\frac{t}{\eta}\right)^\beta\right), \quad h(t) = \frac{\beta}{\eta} \left(\frac{t}{\eta}\right)^{\beta - 1}$$

• CO Detectors: $\eta = 7.0\text{ yr}, \beta = 3.0$ (electrochemical electrolyte dry-out)
• Smoke Detectors: $\eta = 10.0\text{ yr}, \beta = 2.5$ (optical chamber degradation)
• HVAC Furnaces: $\eta = 15.0\text{ yr}, \beta = 2.2$ (heat exchanger thermal fatigue)
• Water Heaters: $\eta = 10.0\text{ yr}, \beta = 2.0$ (sacrificial anode depletion)

C. Lumped-Capacitance Pipe Freeze Model (proactive/engine.py)

Before issuing freeze warnings, the Proactive Engine evaluates unheated pipe-cavity cooling ($Q_{\text{furnace}} = 0$):

$$T_{\text{zone}}(t) = T_{\text{out}} + \big(T_{\text{zone}}(0) - T_{\text{out}}\big) \cdot e^{-t / \tau_{\text{pipe}}}$$

• Deterministic Trajectory ($\tau_{\text{pipe}} = 16.0\text{ h}$, $T_{\text{out}} = 24^\circ\text{F}$): Starting from $68^\circ\text{F}$ indoor setpoint, the pipe zone reaches $32^\circ\text{F}$ in $t = -16.0 \cdot \ln(8/44) = \mathbf{27.3\text{ hours}}$ (or $21.7\text{ hours}$ if the uninsulated cavity starts cooler at $55^\circ\text{F}$; whole-house envelope $\tau = 35\text{ h}$ reaches $32^\circ\text{F}$ in $59.7\text{ hours}$).
• Conditioned Monte Carlo ($N = 5,000$, seed 42): Under $\tau_{\text{pipe}} \sim \text{Lognormal}(\ln(16), 0.3)$ and $T_{\text{out}} \sim \mathcal{N}(24^\circ\text{F}, 2.5^\circ\text{F})$, $P(T_{\text{pipe}} \le 32^\circ\text{F} \text{ within } 48\text{ h}) = \mathbf{94.1%}$ ($P_{10} = 18.7\text{ h}$, $P_{50} = 27.1\text{ h}$; sensitivity across alternative priors: $\tau = 24\text{ h} \rightarrow 66.0%$, $\tau = 35\text{ h} \rightarrow 24.6%$).

---

3. Transparent Benchmarks & Automated Test Suite (22/22 Passing)

Run python benchmarks/run_benchmarks.py (writes receipt to benchmarks/latest_benchmark_receipt.json) and pytest -v tests/ to reproduce all metrics:

Benchmark / Verification Gate	Measured Result	Honest Scope & Method
20-Device Synthetic Cadence Suite	95.0% (19/20)	Multi-burst synthetic WAV regression across all 20 beepdb rows (clip_04 Kidde smoke 3,100 Hz vs Kidde CO 3,200 Hz at 30s cadence resolves to higher critical severity).
Physical Smoke Piezos (Bloofrzo CC0)	2/2 Calibrated	Real 1.0s single-chirp recordings (#819808 at 3,368.1 Hz, #819807 at 3,122.3 Hz); used to calibrate the ±350 Hz band with user-supplied 30s/60s cadence.
Real Hardware Tone DSP Extraction	4/4 Verified	Real PC POST beep, microwave beep, buzzer, and alarm recordings verified for accurate Hilbert pulse width and Hann FFT peak extraction.
ESC-50 Domestic Noise Smoke Test	6/6 Spec · 6/6 Peak	n=12 smoke test on 6 real ESC-50 domestic clips (washer, vacuum, water, knock, steps, clock): 6/6 standalone clips stay < 0.50 confidence; 6/6 +15 dB chirp mixes recover the 3,368 Hz peak via Hann FFT vs 1/6 for naive global zero-crossing.
Synthetic + ESC-50 Noise Bed	90.0% (9/10)	10 synthetic cadence clips mixed with real ESC-50 domestic room noise at +15 dB and +6 dB SNR.
Local DSP + Scrub + SQLite Pipeline	13.52 ms	9.80 ms FFT/Hilbert + 0.005 ms data.fill(0.0) + 0.016 ms beepdb lookup + 3.70 ms SQLite write.
Automated Pytest Suite	22/22 Passing	Includes test_mcp_streamable_http_spec_handshake, test_simulator_mcp_client_and_cors_session_header, and test_audio_never_persisted_to_disk_or_db.

pytest -v tests/
python benchmarks/run_benchmarks.py

---

4. Quickstart & Cloud Deployment

Option 1: Local Python Environment (Full Live MCP Server + Simulator)

python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -r requirements.txt

# Launch FastAPI + MCP Streamable HTTP Server
python -m uvicorn server.main:app --host 127.0.0.1 --port 8000

Open http://127.0.0.1:8000/client/index.html. The simulator header will display a green Live MCP Session (...) · Spec 2025-11-25 badge confirming an active JSON-RPC 2.0 session over POST /mcp.

Option 2: Docker Compose

docker compose up --build

Option 3: Static AWS S3 Simulator Preview

• URL: http://genius-simulator-115207986956.s3-website.eu-west-2.amazonaws.com
• Displays Offline Replay (No MCP Server Connected) until pointed at a running MCP server via the ⚙️ Server URL button.

---

5. Architectural Decision Records (ADRs)

• ADR 001 (Proactive Push Transport): Uses Server-Sent Events (/events) to push unprompted notifications to the simulator alongside a dormant MCP notification stub (proactive/transports/mcp_push.py).
• ADR 002 (Smart Home API Deferral): Defers legacy OAuth Smart Home skill linking in favor of self-hosted MCP Streamable HTTP (2025-11-25).
• ADR 003 (Simulator-Only Demo Path): Because Alexa+ builder tooling is gated, the demo runs on the labeled "ALEXA+ PREVIEW DEVICE SIMULATOR" (client/app.js), which acts as a real MCP Streamable HTTP client calling POST /mcp. Echo hardware integration is not claimed.

---

6. Licensing & Attributions

• Code: MIT License (both genius and standalone beepdb).
• Data & Audio Corpus: CC-BY 4.0 and CC0 Public Domain. Full attributions in ATTRIBUTIONS.md.