Skip to content
Carlos Rubio-Marroquin
All projects

Agentic Data Analyst

Live

Ask a question about Seattle's building-energy data in plain English. An AI agent works out what to look up, writes the database query itself, runs it, and explains what came back.

It is built directly on the Claude API, with no agent framework and no canned queries. Every answer comes from the City of Seattle's public building energy benchmarking data.

Where the data comes from

Every number on this page comes from one public dataset: Seattle Building Energy Benchmarking, published by the City of Seattle on its open data portal. None of it is made up, sampled or hand-picked.

One row is one building, one year
City law makes every commercial and large apartment building over 20,000 sq ft report how much energy it used and how much it emitted, once a year. Each of those yearly reports is one row. (Seattle ordinance SMC 22.920.)
Eleven years of it
2015 through 2025. More buildings joined the programme over that time: about 3,175 reported in 2015, about 3,700 by 2024. So when a yearly count rises, part of that is simply more buildings reporting.
A fixed copy, not a live feed
The agent reads its own copy of the data rather than calling the city on every question, so answers come back fast and the same question always gives the same result. Seattle publishes new figures about once a year, and updating the demo means loading those in and redeploying it.

So the questions it can answer look like these

  • CompareWhich property types have the worst emissions intensity?
  • TrendHow has office energy use changed since 2018?
  • RankWhich buildings emitted the most in 2023?
  • PredictWhat would the model expect for a 40-year-old, 200,000 sq ft office?

What it can't do: anything the reports don't record. There is nothing here on tenants, rents or occupancy, nothing after 2025, and no other cities. It can tell you what was reported, but not why a number moved.

How it works

There is no query template and no fixed pipeline. The model is given five tools and decides for itself which to call, in what order, and when it has enough to answer.

  1. 01
    list_datasetsdescribe_schema

    Looks up what is in the data

    Before writing anything, it asks which tables and columns actually exist. That way it never guesses at a field that isn't there.

  2. 02
    run_sql

    Writes and runs its own query

    It writes the database query itself and runs it. The connection is read-only, and a guardrail limits how much any one query can pull back.

  3. 03
    predict

    Calls a trained model

    Ask what a building would emit and it hands the question to a model trained on this same data, rather than guessing. The model scores 0.86 on buildings it has seen before and 0.76 on ones it hasn't (ROC-AUC), and it passes that second number along with every prediction.

  4. 04
    make_chart

    Draws a chart when a chart is the answer

    When the answer is a trend or a comparison rather than a single number, it decides that itself and draws the chart.

And when it gets it wrong: a query that fails or comes back empty goes straight back to the model, which reads the error and writes a new one. That self-correction is the difference between an agent and a single prompt.

annual filings
38,309
buildings
3,871
covered
2015–2025
per question
~$0.004

8 questions it has already answered

Real questions the agent has already answered, captured ahead of time. Pick one to see the answer, the chart, and the SQL it wrote. These are saved with the page, so they load whether or not the live agent is awake.

Question

In 2023, which property types had the highest median GHG emissions intensity per square foot? Only consider types with at least 20 buildings.


Answer

In 2023, Supermarket/Grocery Stores had by far the highest median GHG emissions intensity at 5.18 kg CO2e/sf, followed by Laboratories at 4.95 kg CO2e/sf. These two property types stand out as substantially higher than all others.

Here are the top 10:

Property TypeBuilding CountMedian GHG Intensity (kg CO2e/sf)
Supermarket/Grocery Store355.18
Laboratory324.95
Other - Recreation223.62
Hotel781.75
Senior Living Community401.51
K-12 School1451.21
Worship Facility611.19
College/University251.09
Mixed Use Property1411.06
Other370.79

The most common property types—Multifamily Housing (1,881 buildings) and Office (454)—are among the cleanest per square foot, with medians of 0.40 and 0.38 kg CO2e/sf respectively. The intensity variation across property types is over 46-fold, reflecting fundamental differences in operations: supermarkets and labs are energy-intensive; warehouses and multifamily are relatively efficient.

The SQL it wrote
SELECT b.primary_property_type,
       COUNT(*) AS n,
       median(r.ghg_emissions_intensity) AS median_kgco2e_sf
FROM energy_records r
JOIN buildings b USING (ose_building_id)
WHERE r.data_year = 2023 AND r.ghg_emissions_intensity IS NOT NULL
GROUP BY 1
HAVING n >= 20
ORDER BY median_kgco2e_sf DESC;

Question

How has the average weather-normalized site EUI for office buildings changed over time? Include a chart.


Answer

Office buildings in Seattle have achieved significant energy efficiency gains over the past decade. The average weather-normalized site EUI declined from approximately 62.6 kBtu/sf in 2015 to 44.8 kBtu/sf in 2025—a reduction of nearly 28% over the period.

The downward trend has been especially pronounced since 2020, with a sharp drop from 57.2 kBtu/sf in 2019 to 50.0 kBtu/sf in 2020. This likely reflects the shift to remote and hybrid work patterns. Since then, the efficiency has continued to improve gradually, reaching the lowest level on record in 2025 at 44.8 kBtu/sf.

The year-over-year improvements show consistent progress, with only a minor uptick in 2022 (50.5 vs. 49.5 kBtu/sf in 2021) before resuming the downward trajectory.

Average Weather-Normalized Site EUI for Office Buildings (2015–2025)
40 45 50 55 60 65 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 44.8 Data year Average site EUI (weather-normalized)
Show the numbers
Data yearAverage site EUI (weather-normalized)
201562.6
201661.8
201760.5
201858.4
201957.2
202050
202149.5
202250.5
202348.8
202446.6
202544.8
The SQL it wrote
SELECT r.data_year, AVG(r.site_eui_wn) AS avg_site_eui_wn
FROM energy_records r
JOIN buildings b USING (ose_building_id)
WHERE b.primary_property_type = 'Office' AND r.site_eui_wn IS NOT NULL
GROUP BY 1
ORDER BY 1;

Question

What share of annual filings are marked compliant?


Answer

93.4% of annual filings are marked Compliant, with 6.6% marked Not Compliant.

Across the ~38k annual energy records in the dataset:

  • Compliant: 35,796 filings (93.4%)
  • Not Compliant: 2,513 filings (6.6%)
The SQL it wrote
SELECT compliance_status, COUNT(*) AS n,
       ROUND(100.0 * COUNT(*) / SUM(COUNT(*)) OVER (), 1) AS pct
FROM energy_records
WHERE compliance_status IS NOT NULL
GROUP BY 1;

Question

Chart how the number of buildings reporting each year has changed.


Answer

The number of buildings reporting energy data grew steadily from 2015 through 2020, then dipped slightly in 2021 before recovering and plateauing in recent years. Reporting increased from 3,068 buildings in 2015 to a peak of 3,565 in 2023, with 2024 at 3,529 buildings. The 2021 dip (3,137) likely reflects pandemic-related disruptions to compliance filings. Overall, the mandatory reporting program has achieved relatively stable coverage of around 3,500 buildings since 2022.

Number of Buildings Reporting Energy Data by Year
3,000 3,100 3,200 3,300 3,400 3,500 3,600 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 3,332 Data year Number of buildings
Show the numbers
Data yearNumber of buildings
20153,068
20163,132
20173,203
20183,290
20193,356
20203,432
20213,137
20223,431
20233,565
20243,529
20253,332
The SQL it wrote
SELECT data_year, COUNT(DISTINCT ose_building_id) AS num_buildings
FROM energy_records
WHERE ghg_emissions_intensity IS NOT NULL
GROUP BY data_year
ORDER BY data_year

Question

Which neighborhood has the most benchmarked buildings?


Answer

The EAST neighborhood has the most benchmarked buildings, with 576 buildings in the dataset.

The SQL it wrote
SELECT b.neighborhood, COUNT(DISTINCT b.ose_building_id) AS building_count
FROM buildings b
WHERE b.neighborhood IS NOT NULL
GROUP BY b.neighborhood
ORDER BY building_count DESC
LIMIT 1;

Question

What would the model expect for a 40-year-old office building of about 200,000 square feet with no ENERGY STAR score?


Answer

The model predicts this building is unlikely to be a high emitter — it scores a 44% probability of exceeding the median GHG intensity for offices in its year.

This makes sense: while the 40-year age works against it (older buildings tend to be less efficient), the large floor area (200,000 sq ft = log₁₀ ~5.30) and the absence of an ENERGY STAR score (which typically means the building hasn't been formally assessed or certified) don't push it into the high-emitter category. The model assigns it a moderate risk profile.

Note: The model generalises to genuinely new buildings at ~0.76 ROC-AUC, so these predictions are directional rather than definitive. If you can add details like number of floors, parking share, or council district, the prediction would be more refined.

Question

What is the median ENERGY STAR score for Multifamily Housing buildings in 2023?


Answer

The median ENERGY STAR score for Multifamily Housing buildings in 2023 is 88, indicating that these buildings are performing well above the national median of 50.

The SQL it wrote
SELECT median(r.energy_star_score) AS median_energy_star_score
FROM energy_records r
JOIN buildings b USING (ose_building_id)
WHERE b.primary_property_type = 'Multifamily Housing' 
  AND r.data_year = 2023
  AND r.energy_star_score IS NOT NULL

Question

Which buildings had the highest total GHG emissions in 2023?


Answer

The top emitters in 2023 by absolute total GHG emissions are dominated by large institutional buildings and campuses, led by nursing homes, universities, and hospitals:

BuildingTypeTotal GHG Emissions (metric tons CO2e)
Sea-Mar Community Care Center (Nursing Home)Other126,805
UW Seattle Campus (Master)College/University119,701
Boeing Plant 2 - CampusMixed Use20,400
Waste Management Maintenance Shed, Bldg AOther - Services10,188
UW Medical CenterHospital9,527
Seattle Children's Hospital (Main Campus)Hospital7,360

The nursing home and UW campus dwarf the other buildings, with emissions over 5–6 times higher than the third-place Boeing facility. This is expected because total emissions scale with both size and energy intensity—large institutions serving continuous operations (24/7 facilities like hospitals and residential care) use far more energy than typical office buildings. The rankings are dominated by campuses, hospitals, hotels, and other large or energy-intensive operations.

The SQL it wrote
SELECT b.building_name,
       b.address,
       b.primary_property_type,
       r.total_ghg_emissions,
       r.data_year
FROM energy_records r
JOIN buildings b USING (ose_building_id)
WHERE r.data_year = 2023 AND r.total_ghg_emissions IS NOT NULL
ORDER BY r.total_ghg_emissions DESC
LIMIT 20;

Ask your own

This one is live. Your question goes to the real agent, which will plan, query, and answer in front of you, usually in 10 to 30 seconds. There's a limit per visitor so the demo can't run up a bill.

Try