1. Dynamic account selection
New processed buildings can appear as new demo accounts in the selector without a frontend code change.
← All projects
03 / ENERGY / DATA ENGINEERING
Energy Customer Insights
Explain electricity-usage changes through governed data, weather context, automated onboarding, and AI-assisted insight.
Working prototype · automated building onboarding validated
View public GitHub repository →Overview
Explain the change, not just the number.
Energy Customer Insights transforms modeled residential energy data into account-level analytics, month-over-month investigation, weather-aware context, and plain-language explanation. The analytics pipeline computes the facts; the LLM's job is to describe those already-computed results without inventing a cause.
The problem
A higher usage number does not explain whether the main driver was cooling, heating, seasonality, or a change in daily consumption. The application brings account context, trends, end-use composition, and aligned weather together before generating an explanation.
My contribution
Built the AWS S3 ingestion pattern, scheduled Lambda transfer, Databricks Auto Loader and Bronze/Silver/Gold layers, automated building onboarding, weather enrichment, dynamic account creation, MLflow serving, frontend discovery, and evidence-constrained AI explanation workflow.
Architecture
Automated from source file to frontend account.
S3 incomingNew modeled building source file
AWS LambdaScheduled copy / verify / archive
Databricks ingestionAuto Loader → Bronze → Silver
Gold analyticsDaily / monthly / investigation + weather
Model servingDeterministic evidence + AI explanation
FrontendDynamic account discovery, charts, investigation
The latest validation added a fourth building and confirmed that it flowed through S3 transfer, lakehouse processing, weather enrichment, Gold, snapshot refresh, live serving, and frontend discovery without manual application changes.
Application walkthrough
Account discovery → trends → investigation → explanation.
The account list is dynamic; the production frontend does not hardcode how many demo accounts exist.
2. Account overview and trends
Annual usage, daily usage, peak demand, building context, monthly/daily charts, and end-use composition are presented together.
3. Investigation and AI explanation
Month-over-month usage, end-use contribution, degree days, and temperature context are computed first; the explanation summarizes that evidence.
Key capabilities
- Automated source transfer: Lambda uses a copy-verify-then-delete pattern between S3 incoming, landing, archive, and audit paths.
- Automated building onboarding: Silver metadata matching, weather mapping, Gold account summary, snapshot refresh, and frontend discovery are data-driven.
- Evidence-first AI: monthly comparison, end-use deltas, and degree-day metrics are computed before the LLM is called.
- Resilient explanation: if the LLM path fails, the application can fall back to a deterministic non-AI summary and label it accordingly.
Representative investigation: DEMO-102517, May 2018
Usage increased from 258.51 kWh in April to 584.69 kWh in May. The net increase was 326.18 kWh, average daily usage rose 118.88%, and 98.93% of the net increase was associated with cooling in the demonstration data.
4 accountsGold, snapshot, live endpoint, and frontend account count after the validated fourth-building run
No manual UI changeThe new building appeared automatically in the frontend account selector
98.93%Cooling share of the representative May net usage increase
Modeled ResStock demonstration data; not a production utility billing or customer-service system.