Indeklima Twin

Every room in a building is designed with a promise. Indeklima Twin checks whether it keeps it.

▶ Live site + interactive 3D twin · 60-second film · PDF report

Simulated sensor data. The feed comes from a physics model with planted faults, so every diagnosis can be scored against ground truth. The pipeline reads MQTT / JSONL / CSV and takes a real feed unchanged.

What it does

The MEP engineer designed every room for a number of people and an outdoor-air rate. That gives each room a design promise: the CO₂ level and comfort category it should never leave. After handover, nobody checks it. Sensors are compared with generic thresholds ("CO₂ above 1,000 ppm"), so a 15-person meeting room and a corridor are judged by the same line.

Indeklima Twin connects the two things the building already has:

flowchart LR
    A["🏗️ IFC model<br/>IfcSpace geometry"] --> P{{"Design promise<br/>per room"}}
    B["📐 MEP design<br/>EN 16798-1 cat. II"] --> P
    S["📡 Sensor feed<br/>CO₂ · T · RH · VAV · people"] --> Q["Ingest + QC<br/>DuckDB → Parquet"]
    Q --> X["Analytics<br/>categories · overheating<br/>airflow from CO₂ decay"]
    P --> X
    X --> D["Diagnosis<br/>5 rule families<br/>with evidence"]
    D --> W1["IFC4 Psets"]
    D --> W2["BCF 2.1 issues"]
    D --> W3["Brick graph"]
    D --> W4["3D dashboard"]
    D --> W5["MCP: ask in<br/>plain language"]
    style P fill:#1f6fd6,color:#fff,stroke:#1f6fd6
    style D fill:#d63b3b,color:#fff,stroke:#d63b3b

Related MCP server: Keystone MCP Server

The live twin

The 3D dashboard is built straight from the IFC spaces: exploded storeys, rooms coloured by CO₂, temperature, airflow-vs-design or occupancy, a time slider over two weeks, and a promise card for every room.

Open the interactive twin →

Promise vs reality

35 of 37 rooms keep the winter CO₂ promise. The two that don't break it for very different reasons. Room 2.03 is really under-ventilated: its VAV damper is stuck. Room 2.11 only looks like it: its CO₂ sensor is drifting, and the VAV is over-ventilating on the false reading. A fixed-threshold alarm can't tell them apart. The twin can.

Diagnoses, not alarms

Ask your building (MCP)

src/mcp_server.py is a Model Context Protocol server. Any MCP client (Claude, Copilot, IDE agents) can query the twin in plain language through seven tools: building_overview, list_rooms, room_promise, list_findings, room_timeseries, query (sandboxed read-only SQL) and export_bcf.

The GIF is a real recorded session (results/mcp_claude_demo_full.json). Claude called building_overview → list_findings → list_rooms → room_promise, separated the real breach from the drifting sensor, proposed an order of work, and flagged on its own that the data is simulated.

{"mcpServers": {"indeklima-twin": {"command": "python", "args": ["<repo>/src/mcp_server.py"]}}}

SQL runs on an in-memory copy with external access disabled and configuration locked. tests/test_mcp.py drives every tool through the official MCP client and checks that writes and file reads are refused.

Write-back: nothing new to learn

Proof

The detectors never see the simulator's truth.

Two detectors failed during validation and were fixed for physical reasons, not by tuning thresholds. The shading rule moved from peak temperature (the VAV hid the fault by cooling harder) to an inverse energy balance. The setback rule moved from "warmer than the setpoint" (heavy rooms stay warm overnight) to "warmer than the other rooms on the same night".

Where it fits

Simulation tools predict the promise. This checks whether it's kept. See docs/POSITIONING.md for the full comparison and the honest gaps.

Run it

pip install -r requirements.txt
# config/local_paths.json -> {"source_ifc": "<path to your IFC>"}   (see config/local_paths.example.json)
python src/ifc_extract.py        # IFC -> data/rooms.json + web/building.glb
python src/simulate.py           # synthetic feed; skip it and drop a real feed in data/raw/*.jsonl
python src/run_all.py            # ingest -> QC -> comfort -> CO2-decay airflow -> diagnosis -> results/
python src/validate_ach.py && python src/export_web.py && python src/twin/writeback.py
python -m pytest tests           # acceptance + MCP tests
python src/robustness.py         # 16-seed validation incl. held-out seeds
python src/mcp_server.py         # MCP server over stdio

data/rooms.json (the room registry extracted from the IFC) ships with the repo, so everything after the first step runs without the source model.

Repository

src/             pipeline: IFC extract, simulator, ingest, QC, comfort, ventilation, diagnosis, write-back, MCP server
src/twin/        the analysis package (runs unchanged on real data)
tests/           acceptance tests + end-to-end MCP test
web/             3D dashboard (Three.js) and its data
docs/            GitHub Pages site, media, positioning, case study
results/         validation outputs (robustness, airflow, findings, MCP session)
reports/         PDF report

Geometry: IFC4 renovation design model of Building 308. Design airflows: DS/EN 16798-1 cat. II (0.7 L/s·m² + 7 L/s·person). Weather: Copenhagen-Kastrup TMYx (climate.onebuilding.org). Sensor data are simulated.