Carrier · AI Solutions Architect · Round 1 · Architecture Deep Dive (Supplemental)

Interview: 2026-05-19 morning · Jai (Jaivenkat Krishna H) · AI-CoE Lead · 30-person team
Use alongside Front/Back sheets · Source: Multi-agent SOTA research 2026-05-18 · Primary sourced
Central finding from SOTA research: Carrier's AWS+GCP dual-cloud pattern is the Snowflake Cortex model applied twice — proprietary AI embedded inside the data platform, data stays in-perimeter, intelligence runs in-place. GCP was chosen because WeatherNext (DeepMind) doesn't exist on AWS. Follow the proprietary embedded AI to its native platform. That's the entire logic.

The Cortex Analogy — Speak It Exactly TOGAF LANGUAGE

Snowflake Cortex = intelligence embedded inside the data platform. Data stays inside the Snowflake perimeter. AI runs against it in place — not from an external AI service.

The Pattern — Three implementations:

SNOWFLAKE → Data + Cortex LLM inference inside Snowflake perimeter. Cortex Analyst, Search, Intelligence are native capabilities — no external AI call.
AWS → Abound: IoT telemetry ingests into IoT Core, stays in AWS perimeter. Bedrock + SageMaker run in-place. Same Cortex logic: platform-embedded AI against data that doesn't leave.
GCP → HEMS: Energy/grid data into BigQuery, stays in GCP perimeter. WeatherNext (DeepMind) runs in-place as a native GCP service. Identical pattern.

Where it breaks: No unified cross-cloud semantic layer spans AWS+GCP. Two separate Cortex-pattern stacks for two separate product lines. Not a single intelligence layer — no public evidence of cross-cloud integration.

GCP Weak-Spot — REVISED ANSWER REPLACES OLD

Previous version: "Dialogflow + Kubeflow → Vertex AI" — weak, generic, doesn't reflect why Carrier is on GCP.

Lead with WeatherNext — the actual reason:
"Carrier is on GCP specifically because WeatherNext — Google DeepMind's AI weather forecasting model — is exclusive to GCP. There's no AWS equivalent built by a research organization at that caliber. For a product like Carrier Energy's HEMS — a heat pump and battery system that needs to pre-condition buildings, charge/discharge storage, and shed grid load based on tomorrow's weather and real-time grid pricing — WeatherNext isn't a preference, it's the enabling technology. That's why GCP. The logic is the Cortex pattern: follow the proprietary embedded AI to its native platform."
Then bridge to your experience:
"For Vertex AI: Kubeflow Pipelines v2 is the native Vertex AI pipeline framework. My Optinosis work on Kubeflow/Kubernetes maps directly — same open-source DAG orchestration, same pipeline abstraction. BigQuery is the analytics gold layer — same medallion pattern as my PostgreSQL architecture at NewsRx, different managed platform. The concepts transfer cleanly."

AWS — Full Abound Stack KNOW THIS COLD

LayerServiceFunction
IngestionAWS IoT CoreBACnet, MQTT, webhooks from building systems
RoutingIoT Rule ActionsCarrier IoT Action Engine (serverless routing)
StreamingKinesis Streams/FirehoseReal-time telemetry processing
ComputeAWS LambdaServerless event handling
Operational DBDynamoDBPer-asset operational data
Time-seriesTimestreamSensor telemetry at scale
AnalyticsRedshiftPortfolio-level reporting, carbon metrics
GraphNeptuneAsset relationship modeling
Data lakeS3Raw telemetry store, ML training data
ETLGlueSilver-layer feature engineering
MLSageMakerFault prediction training + inference
GenAIBedrockNet Zero Mgmt + Tell Me More
OCRTextractMulti-language utility bill parsing
APIAPI GatewayAbound API surface

Scale: 150K+ connected devices · 32K+ sites · 640M+ sq ft · 7B+ kWh lifetime · 40K+ dispatch saves

Workload Partitioning — Not Vendor Hedging

PatternWhat It MeansIs This Carrier?
Vendor hedgingSame workload on both clouds; negotiate leverageNo
Multi-cloud redundancyReplicate for failover; no single cloud failureNo
Workload partitioningDifferent workloads on different clouds; each cloud owns a unique capability the other can't replicateYes — this is it

The partition: Abound = commercial buildings on AWS. HEMS = residential energy on GCP. Product-line separation. No cross-cloud data flow documented. Two independent Cortex-pattern stacks.

Interview framing: "The candidate who says 'vendor diversification' demonstrates surface familiarity. The right answer is workload partitioning driven by platform-embedded AI — each cloud chosen because that AI model only runs there."

GCP Stack — Carrier Energy HEMS

ServiceRole
WeatherNextDeepMind AI weather forecasting — the reason GCP was chosen. Pre-conditioning, battery optimization, grid demand response. No AWS equivalent.
BigQueryData management + analytics. Energy/grid/residential telemetry gold layer.
Vertex AIML training + inference. Maps from Kubeflow (Optinosis experience).
GenAIOptimization layer — not named Gemini in any confirmed press release as of May 2026.

Partnership announced 2025-03-05 — same day WeatherNext enterprise launch. Co-announcement was deliberate.

5 Business Drivers — Why This AI Acceleration Now

#DriverCarrier-Specific Evidence
1 Grid stress / AI data centers QuantumLeap data center HVAC: $1B est. 2025 rev → $1.5B 2026 guide. Q1 2026 data center orders +500% YoY. Buildings as grid assets, not just consumers.
2 Decarbonization regulation EU EPBD, building benchmarking laws force portfolio Scope 1+2 tracking. Bedrock + Textract for multi-language utility bill processing is the product response. 437,900 metric tons CO2e prevented.
3 HVAC technician shortage Tell Me More (Feb 2026) explicitly frames AI-assisted diagnostics as labor substitution. Knowledge transfer to less-experienced technicians at the fault point.
4 Margin expansion: hardware → SaaS Abound on AWS Marketplace: $28K/year base contract. Recurring revenue above the hardware sale. Aftermarket double-digit growth target. SaaS margins structurally higher.
5 Portfolio concentration → concentrated AI investment $10B+ divestitures (fire, security, refrigeration) + €12B Viessmann acquisition = pure-play climate company. Every connected asset feeds one AI system. Scale compounds faster.

CoE stakes framing: Carrier no longer has fire, security, or access as portfolio buffers. The 30-person AI CoE is building the AI depth that defends market position against Honeywell's broader platform. This is high-stakes, not exploratory.

Competitive Landscape CONTEXT FOR JAI

CompanyPrimaryGCP?AI Platform
CarrierAWSYes (WeatherNext)Bedrock + SageMaker + Abound
HoneywellAzureYes (Gemini/Vertex)Forge / Connected Solutions
Johnson ControlsAzureNoOpenBlue + Azure OpenAI
SiemensAzureNoBuilding X + Industrial Copilot
Trane (BrainBox)AzureNoARIA + per-asset ML

Pattern: The two companies running GCP (Carrier, Honeywell) are doing the most architecturally interesting AI work. Both chose GCP specifically for DeepMind-class capabilities unavailable on their primary platform. The Azure-only players have narrower AI differentiation.

Carrier's structural vulnerability: Sold fire/security/access. Cannot bid for unified smart building contracts requiring a single vendor. Abound AI depth is the only competitive answer — no platform breadth to fall back on.

Architecture Vocabulary — Use These with Jai

Cortex pattern — proprietary AI embedded inside a data platform; data stays in-perimeter; intelligence runs in-place. Carrier does this twice.
Workload partitioning — distinct workloads on distinct clouds by unique capability, not vendor hedging or resilience architecture.
Platform-embedded AI — the forcing function for both cloud choices. WeatherNext is only on GCP. Bedrock is only on AWS. Platform choice follows the model.
Data perimeter — each cloud is a self-contained intelligence perimeter. No cross-cloud data flow. Two perimeters, two product lines, no integration layer needed.
Non-intrusive inference — Jai's NILM patent language. Use it. "The Abound per-asset pattern — inferring equipment state from load signatures without per-component sensors — is the NILM architecture. My Invistics work is the same pattern applied to drug diversion detection."