BuildOcc

ATUS-grounded LLM occupant agents for building energy simulation.

Agents are initialized from American Time Use Survey (ATUS) population microdata, reason over environment state using an LLM at each 15-minute timestep, and accumulate a memory stream that enables persistent behavior change. The library is exposed through a three-layer platform interface — Python library, REST API, and MCP server — so any building energy tool can integrate an occupant behavioral layer without bespoke coupling code.

Architecture

┌─────────────────────────────────────────┐
│  Layer 3 — MCP Server                   │  ← Claude / LLM orchestrators,
│  (tool-calling interface for LLM apps)  │     Home Assistant (native MCP client)
└────────────────┬────────────────────────┘
                 │ wraps
┌────────────────▼────────────────────────┐
│  Layer 2 — REST API (FastAPI)           │  ← EnergyPlus Python callbacks,
│  (standard HTTP, platform-agnostic)     │     VOLTTRON, OpenStudio, any script
└────────────────┬────────────────────────┘
                 │ calls
┌────────────────▼────────────────────────┐
│  Layer 1 — Python Agent Library         │  ← Direct import, unit tests,
│  (core logic, no network dependency)    │     research scripts
└─────────────────────────────────────────┘

Requirements

• Python ≥ 3.11
• An LLM API key — set one of:
  • ANTHROPIC_API_KEY (default provider, claude-haiku-4-5)
  • OPENAI_API_KEY (provider openai, gpt-4o-mini)
  • GOOGLE_API_KEY (provider google, gemini-2.0-flash) — install: pip install "buildocc[google]"
  • Ollama running locally (provider ollama, llama3.2) — no extra package needed

Installation

git clone https://github.com/humanbuildingsynergy/BuildOcc
cd BuildOcc
pip install -e ".[dev]"

# Set your API key — the library loads .env automatically via python-dotenv
cp .env.example .env
# Then edit .env and replace the placeholder with your real key

Quick Start

Layer 1 — Python library (direct import)

from datetime import datetime
from occupant_agent import OccupantAgent, DeviceState, EnvironmentState, RoomState, ActivityScheduler

# Create an ATUS-grounded agent for an employed single adult (25–44)
agent = OccupantAgent.from_stratum("O1", seed=42, llm_provider="anthropic")

# Sample the agent's current activity from empirical ATUS distributions
scheduler = ActivityScheduler(stratum="O1", seed=42)
timestep = datetime(2024, 7, 15, 19, 0)
atus_code = scheduler.sample(timestep)  # e.g. "120303" — computer leisure

# Build an environment state
env = EnvironmentState(
    timestep=timestep,
    zone_temp_c=25.0,
    outdoor_temp_c=34.0,
    tou_rate=0.22,  # peak rate $/kWh
    devices=[
        DeviceState(device_id="hvac",   state=True,  power_w=3500),
        DeviceState(device_id="washer", state=False, power_w=500),
    ],
    rooms=[RoomState(room_id="living_room", occupied=True)],
)

# Step: agent reasons over persona + memory + environment → action
action = agent.step(env, atus_code=atus_code)
print(action.action_type, action.target_id, action.reasoning)
# e.g. "toggle_device" "hvac" "Peak rate is $0.22/kWh and I'm comfortable..."

# Send a demand-response signal (Type B — educational)
response = agent.receive_signal(
    signal_type="B",
    content="Your HVAC costs 3× more before 9pm. Raising the setpoint by 2°C saves ~$0.35 today.",
    env=env,
)
print(response.response, response.reasoning)
# e.g. "accepted" "Makes economic sense and I'm not uncomfortable at 26°C..."

Signal types:

Type	Label	Description
A	Direct command	"Turn off the dishwasher until 9pm"
B	Competence-building (boost)	Price/cost explanation for why a change helps
C	Social norm (nudge)	Comparison to similar households

Demographic strata:

ID	ATUS Stratum
O1	Employed adult, single, 25–44
O2	Retired couple, 65+
O3	Employed parent with children, 35–54
O4	Unemployed adult, 25–44

Full simulation loop

The library reads ANTHROPIC_API_KEY (or whichever provider key) directly from .env — no shell export needed.

# Runs 8 timesteps of a O1 evening with ATUS-sampled activities + a Type B signal
python3 examples/simulation_loop.py --stratum O1 --seed 42 --steps 8

# Other providers
python3 examples/simulation_loop.py --stratum O2 --provider openai --steps 16 --start-hour 6
python3 examples/simulation_loop.py --provider google --stratum O3   # Google Gemini
python3 examples/simulation_loop.py --provider ollama --stratum O1   # local Ollama

# Offline / no API key (deterministic mock responses — useful for CI and testing)
python3 examples/simulation_loop.py --mock --stratum O1 --steps 4

# Other flags
python3 examples/simulation_loop.py --zone-temp 78.5   # override zone temp (°C)
python3 examples/simulation_loop.py --hardcode          # fixed ATUS codes (ablation baseline)

Demand response signal demo

Shows all three signal types (A/B/C), cross-stratum comparison, and the extra_context kwarg:

python3 examples/signal_demo.py                         # all three parts, Anthropic
python3 examples/signal_demo.py --part 2                # stratum comparison only
python3 examples/signal_demo.py --provider openai --warmup 4
python3 examples/signal_demo.py --mock                  # offline mock mode

Layer 2 — REST API

# Start the server
ANTHROPIC_API_KEY=... buildocc-api
# or: uvicorn occupant_agent.api.app:app --reload --port 8000

# Initialize an agent
curl -s -X POST http://localhost:8000/agents/initialize \
  -H "Content-Type: application/json" \
  -d '{"stratum": "O1", "seed": 42}' | python3 -m json.tool
# → {"agent_id": "...", "stratum": "O1", "seed": 42}

# Step (replace AGENT_ID)
curl -s -X POST http://localhost:8000/agents/AGENT_ID/step \
  -H "Content-Type: application/json" \
  -d '{
    "environment": {
      "timestep": "2024-07-15T19:00:00",
      "zone_temp_c": 25.0, "outdoor_temp_c": 34.0, "tou_rate": 0.22,
      "devices": [{"device_id": "hvac", "state": true, "power_w": 3500}],
      "rooms": [{"room_id": "living_room", "occupied": true}]
    },
    "atus_code": "120303"
  }' | python3 -m json.tool

# Send a demand-response signal
curl -s -X POST http://localhost:8000/agents/AGENT_ID/signal \
  -H "Content-Type: application/json" \
  -d '{
    "signal_type": "B",
    "content": "Your HVAC costs 3x more before 9pm. Raising the setpoint 2°C saves ~$0.35 today.",
    "environment": {
      "timestep": "2024-07-15T17:00:00",
      "zone_temp_c": 25.0,
      "outdoor_temp_c": 34.0,
      "tou_rate": 0.22,
      "devices": [{"device_id": "hvac", "state": true, "power_w": 3500}],
      "rooms": [{"room_id": "living_room", "occupied": true}]
    }
  }' | python3 -m json.tool

# Check agent state (memory count, last action, reflection history)
curl -s http://localhost:8000/agents/AGENT_ID/state | python3 -m json.tool

REST endpoints:

Method	Endpoint	Description
POST	/agents/initialize	Create agent; returns agent_id
POST	/agents/{id}/step	Advance one 15-min timestep; returns AgentAction
POST	/agents/{id}/signal	Deliver A/B/C signal; returns SignalResponse
GET	/agents/{id}/state	Memory count, last action, reflection history
GET	/agents/	List all agents
DELETE	/agents/{id}	Delete agent and all records
GET	/health	{status: "ok", version: "1.0.0"}

Layer 3 — MCP server

# Requires the REST API to be running first
BUILDOCC_API_URL=http://localhost:8000 buildocc-mcp
# or: python3 -m occupant_agent.mcp_server.server

MCP tools: initialize_agent, step, send_signal, get_state, reset_agent.

Configure in Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "buildocc": {
      "command": "buildocc-mcp",
      "env": {
        "BUILDOCC_API_URL": "http://localhost:8000"
      }
    }
  }
}

Start the REST API first (buildocc-api), then restart Claude Desktop. The initialize_agent, step, and send_signal tools will appear in Claude's tool list.

Testing

# Unit tests (no API key needed — uses MockLLMAgent)
pytest tests/unit/

# Integration tests (no API key needed — uses MockLLMAgent end-to-end)
pytest tests/integration/

# Offline smoke-test of the simulation loop (--mock bypasses the LLM entirely)
python3 examples/simulation_loop.py --mock --stratum O1 --steps 4
python3 examples/signal_demo.py --mock --part 1

How it works

ATUS grounding

Agents are initialized from ATUS 2022+2023 microdata (16,684 respondents; 299,513 activity episodes; 6,611 of these fall in the four modeled strata) matched by demographic stratum. Activity scheduling uses time-at-activity distributions — the population-weighted fraction of respondents in each activity at each clock hour, computed via episode-overlap rather than start-time sampling. This avoids the start-time bias that over-represents long-duration activities (e.g., sleeping). Nine categories are modeled: sleeping, work, food prep, laundry, TV, eating, exercise, travel, and other. Weekday and weekend distributions are computed separately (TUDIARYDAY) and the scheduler selects the appropriate distribution from timestep.weekday().

LLM reasoning at each timestep

step() synthesizes six inputs: the agent's persona (core memory), top-5 retrieved memories by recency + importance score, current environment state (temperatures, TOU rate, device states), current ATUS activity, time of day, and a per-day WFH flag (sampled from persona.sample_wfh_today() and resampled on date rollover). The LLM returns a structured action and a self-rated importance score (1–10) for the observation. Valid target_id values (device IDs and room IDs) are injected into the prompt schema so the LLM cannot hallucinate invalid targets.

Persona diversity

Each create_persona() call samples a demographically grounded agent: income bracket drives comfort_band_c (°C deviation from setpoint before acting — wider for lower income, per RECS 2020), appliance ownership (scaled ±30% around RECS base priors by income position within stratum), and signal preference framing in the LLM prompt (dollar-savings vs. social-comparison language). Four LLM providers are supported for cost/privacy trade-offs: Anthropic, OpenAI, Google Gemini, and local Ollama.

Memory stream and reflection

Follows Park et al. (2023). Retrieval score = 0.5 × recency + 0.5 × (importance/10) with exponential recency decay (24-hour half-life). When the cumulative importance accumulator exceeds a threshold (default 100), reflection fires: the LLM synthesizes the last 30 memories into 3 high-level insights, stored as permanent memory entries. A more capable model is used for reflection; a cheaper model handles routine step() calls.

Persistence

All agent state (persona, memory stream, action log, signal log) is persisted to SQLite via SQLAlchemy. Each API request loads fresh from the database and writes back — stateless HTTP with durable state. Multiple agents can run concurrently via agent_id (UUID4).

How to extend

BuildOcc is designed as a community platform. You can add new demographic profiles, activity schedulers, and memory architectures without forking the repository.

Define a new demographic profile (stratum)

Subclass BasePersona and register it under a key. Your class will be discoverable via OccupantAgent.from_stratum() and list_strata().

import random
from occupant_agent import OccupantAgent
from occupant_agent.core import BasePersona, register_stratum, list_strata

@register_stratum("P5")
class LowIncomeElderlyAlone(BasePersona):
    """Single elderly adult, low income — for low-income housing DR research."""

    def __init__(self, seed=None, **kwargs):
        self._age = random.Random(seed).randint(65, 80)

    @property
    def stratum(self): return "P5"
    @property
    def age(self): return self._age
    @property
    def sex(self): return "female"
    @property
    def income_bracket(self): return 3          # low income (HEFAMINC 1–16)
    @property
    def work_from_home(self): return False
    @property
    def home_gym(self): return False
    @property
    def wfh_probability(self): return 0.0
    @property
    def comfort_band_c(self): return 2.2        # cost-sensitive → wide tolerance
    @property
    def appliances(self): return {"hvac", "thermostat", "tv", "refrigerator"}
    @property
    def schedule_priors(self): return {}
    @property
    def core_memory_text(self):
        return (f"I am a {self._age}-year-old woman living alone on a fixed income. "
                "I keep my thermostat low to save money.")
    def sample_wfh_today(self, rng): return False

# Now usable everywhere:
agent = OccupantAgent.from_stratum("P5", seed=42)
print(list_strata())  # ["O1", "O2", "O3", "O4", "P5"]

Define a new activity scheduler

Subclass BaseScheduler to ground the agent in a different data source (Homer, MTUS, synthetic, etc.).

from datetime import datetime
from occupant_agent.core import BaseScheduler, register_scheduler

@register_scheduler("homer")
class HomerScheduler(BaseScheduler):
    """Activity grounding from Homer dataset (21-participant HAR corpus)."""

    def __init__(self, stratum=None, seed=None, **kwargs):
        ...  # load Homer diary records

    def sample(self, timestep: datetime) -> str:
        ...  # return 6-digit ATUS code

    def category_weights(self, hour: int, timestep=None) -> dict[str, float]:
        ...  # return {category: probability}

# Inject at agent creation:
agent = OccupantAgent.from_stratum("O1", seed=42, scheduler="homer")

Define a new memory architecture

Subclass BaseMemoryStream to replace the retrieval algorithm (e.g., sentence-embedding similarity, graph-structured memory).

from datetime import datetime
from collections.abc import Callable
from occupant_agent.core import BaseMemoryStream

class EmbeddingMemory(BaseMemoryStream):
    """Retrieval weighted by semantic similarity (sentence-transformers)."""

    def retrieve(self, query_time: datetime, k: int = 5):
        ...  # rank by cosine similarity to current context

# Inject at agent construction:
mem = EmbeddingMemory()
agent = OccupantAgent(persona=persona, memory=mem)

Platform-specific extensions via EnvironmentState

Use extensions: dict to pass platform-specific data alongside the core schema without modifying it:

env = EnvironmentState(
    timestep=ts, zone_temp_c=23.0, outdoor_temp_c=29.0, tou_rate=0.18,
    devices=[...], rooms=[...],
    extensions={
        "home_assistant": {
            "sensor.living_room_co2": 650,
            "binary_sensor.front_door": "off",
        },
        "energyplus": {"zone_air_humidity_ratio": 0.009},
    }
)

The core agent ignores extensions; platform-specific plugins can read it.

Package your extension for distribution

# your_package/__init__.py
from occupant_agent.core import register_stratum
from .personas import LowIncomeElderlyAlone, PublicHousingResident

register_stratum("P5")(LowIncomeElderlyAlone)
register_stratum("P6")(PublicHousingResident)

After pip install your-package, any script that imports it gains the new strata.

Built-in extension point summary

Base class	Register with	Discovered by
BasePersona	@register_stratum("P5")	OccupantAgent.from_stratum("P5"), list_strata()
BaseScheduler	@register_scheduler("homer")	OccupantAgent.from_stratum(..., scheduler="homer"), list_schedulers()
BaseMemoryStream	Inject directly into OccupantAgent(memory=...)	—

---

Regenerating ATUS outputs

Pre-generated ATUS outputs are bundled in occupant_agent/data/. Raw ATUS microdata is not committed to the repository — download it and place in data/atus/{year}/extracted/ before running these scripts. To regenerate outputs after a data update:

python3 scripts/atus/analyze.py
# Produces scripts/atus/outputs/time_at_activity.csv       (hour, category, weighted_pct, stratum, day_type)
#          scripts/atus/outputs/activity_frequency_{O1,O2,O3,O4}.csv
#          scripts/atus/outputs/schedule_peak_hours.csv    (weekday/weekend peak hours per stratum)
# Then sync:
cp scripts/atus/outputs/time_at_activity.csv occupant_agent/data/
cp scripts/atus/outputs/schedule_peak_hours.csv occupant_agent/data/

Download ATUS microdata from BLS ATUS (.dat files, uppercase column headers) or IPUMS Time Use (.csv files, lowercase column headers — both formats are supported automatically).

Data

Raw datasets go in data/. See docs/datasets_and_resources.md for access and coverage details.

Directory	Dataset	Purpose
data/atus/	ATUS 2022–23 (BLS)	Agent grounding and scheduling
data/casas/	CASAS Aruba/Milan (Zenodo)	Behavioral validation (Phase 2)
data/ecobee/	DOE ecobee 2017 (OSTI)	Thermostat validation (Phase 2)
data/pecan_street/	Pecan Street Dataport	Energy validation (Phase 2)
data/recs/	EIA RECS	Appliance ownership priors

Project structure

occupant_agent/
    core/
        base_persona.py    — BasePersona ABC (extension contract for new strata)
        base_memory.py     — BaseMemoryStream ABC (extension contract for memory backends)
        base_scheduler.py  — BaseScheduler ABC (extension contract for activity sources)
        registry.py        — Plugin registry: @register_stratum, @register_scheduler
    agent/
        occupant.py     — OccupantAgent: step(), receive_signal(), from_stratum()
        memory.py       — MemoryStream(BaseMemoryStream): retrieval, reflection
        persona.py      — Persona dataclass, create_persona(), ROOM_DEFAULTS
    grounding/
        scheduler.py         — ActivityScheduler(BaseScheduler): ATUS-grounded sampling
        fixed_schedule.py    — FixedScheduleScheduler: rule-based ablation baseline
        activity_code_map.py — ATUS tier-3 code → occupancy + device state mapping
    llm/
        client.py       — call_llm(): Anthropic / OpenAI / Google Gemini / Ollama
    environment/
        state.py        — Pydantic models: EnvironmentState (+ extensions), AgentAction
        simulation.py   — SimulationEnvironment: device/room/setpoint state
    persistence/
        store.py        — AgentStore: SQLite via SQLAlchemy
    analysis/
        simulation_log.py — SimulationLog: per-step records, CSV/JSON export
        metrics.py        — compute_kl, compute_ks, compute_cvrmse, compute_mbe
    testing/
        mock_llm.py     — MockLLMAgent: deterministic test double (no API key needed)
        fixtures.py     — make_env(), make_room() — factory helpers for unit tests
        conformance.py  — assert_persona_contract(), assert_scheduler_contract()
    cli.py              — Entry point for buildocc CLI command
    data/               — Bundled ATUS outputs (time_at_activity.csv, ...)
    api/
        app.py          — FastAPI REST API (Layer 2)
    mcp_server/
        server.py       — MCP server (Layer 3, wraps REST API)
scripts/
    evaluate.py            — Config-driven evaluation harness (KL, KS, action distribution)
    validate_strata.py     — Validate each demographic stratum across a batch of seeds
    validate_signals.py    — Validate Type A/B/C signal response rates per stratum
    extract_zone_temps.py  — Extract zone temperatures from EnergyPlus output CSV
    atus/
        parse.py        — ATUS microdata → per-respondent episode records
        analyze.py      — Episode records → time_at_activity.csv + frequency CSVs
data/atus/              — Raw ATUS 2022+2023 microdata (not committed to Git)
examples/
    simulation_loop.py      — ATUS-grounded evening simulation with Type B signal
    signal_demo.py          — Demand response signal reference: all three types + stratum comparison
    energyplus_callback.py  — EnergyPlus Python API integration pattern
docs/
    datasets_and_resources.md — Dataset access, coverage, and caveats
    methodology_decisions.md  — Design decision log (D1–D13)

Development plan

v1 is complete. Phase 2 (validation against CASAS and Pecan Street behavioral data) is next.

How to cite

If you use BuildOcc in your research, please cite both the software release and the accompanying paper.

Software (this release):

Jung, W. (2026). BuildOcc: ATUS-grounded LLM occupant agents for building energy simulation (v1.0.0). Zenodo. https://doi.org/10.5281/zenodo.21192895

@software{jung2026buildocc,
  author    = {Jung, Wooyoung},
  title     = {BuildOcc: ATUS-grounded LLM occupant agents for building energy simulation},
  version   = {1.0.0},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.21192895},
  url       = {https://doi.org/10.5281/zenodo.21192895}
}

Paper (SoftwareX — publication details to be announced):

@article{jung2026buildoccpaper,
  author  = {Jung, Wooyoung},
  title   = {BuildOcc: A Large Language Model Occupant Agent Platform for Building Energy Research},
  journal = {SoftwareX},
  year    = {2026},
  note    = {To be announced}
}

GitHub's "Cite this repository" button (repo sidebar) generates these automatically from CITATION.cff.

License

Apache License 2.0