HearthRun the simulation

Technical overview

CoStar in, a dispatch plan out.

Every plan starts from a CoStar property record. Vexcel and Matterport fill what the record can’t, a LangChain deepagents surveyor reconciles the three into one sourced profile, and three solvers turn it into crews, routes and a walk to the machine.

Interactive version →
Houston CoStar records
1,758
Houston CoStar records
with a Matterport scan
1,732
with a Matterport scan
Vexcel endpoints
4
Vexcel endpoints
subagents in parallel
3
subagents in parallel

Live

One building, surveyed by the agent

Real CoStar properties from the API. The trace is the last real survey of the building, sped up to about 18 seconds; run it live to start a fresh one.

  1. Events from the deep agent stream here: the plan, three subagents in parallel, their tool calls, conflicts, the profile.

Loading the building…

Pipeline

From record to plan

  1. CoStar extract

    commercial-property.jsonl

  2. normalize_costar

    research market 14 → flat fields

  3. Surveyor deep agent

    costar · aerial · interior

  4. BuildingProfile

    {value, source, confidence}

  5. Planner

    deadline · access · refuge · lot

CoStar

Which record fields Hearth reads, and what each one changes

CoStar fieldHearth fieldWhat it drives
primaryAddress.latitude, longitudelat, lonEvery travel time: straight-line km × 1.4 at 40 km/h
numStoriesAboveGradefloors10 min inside, +4 min per floor climbed without a working elevator
propertyApartment.elevators[], structural.elevators[], isWalkupelevator_countWhether those floors are climbed at all
constructionEvents[eventType=2].year, buildingConstructionMaterialyear_built, constructionHeat time constant → hours until the home is unsafe
propertyType, propertySubtype, apartmentLifestyleTypeproperty_typeHotels and senior housing become refuge candidates
constructionStatus, propertyHospitality.operatingStatusexistingA closed hotel is never offered as a refuge
parking[parkingType=1].totalSpaces, lotSizesurface_spaces, lot_acresA crew staging lot needs 100+ surface spaces
primaryAddress.postcodezipA staging lot sits in a ZIP with no open help request
propertyApartment.models[].units, attachments[].videoKeyfloor_plans, matterport_model_idsWhich Matterport scan to open for a resident's unit

Agents

The surveyor and the refuge finder

  • write_todos plans the run; the task tool fans out to three Haiku subagents in one turn; each saves its findings to /<building_id>/<layer>.md.
  • Every field written needs a source layer and a confidence; a field no layer showed stays unknown. write_profile validates and rejects the rest.
  • CoStar is the authority on elevators and walk-ups. An elevator only seen in a scan is written at confidence ≤ 0.5 and flagged as a conflict.
  • Derived numbers (heat and cold time constants) are computed by code from the fields, never by the model.
  • Refuge finder: CoStar hotels and senior housing outside the dark area, access checked by code, ranked by a deep agent, stopped at a human approval interrupt. Rejections go to the agent’s memory file.

Adjacent data

Vexcel and Matterport calls

● GET /v2/property/extractRoof material and tree cover from the latest capture → heat deadline
● POST /v2/elements/detectDriveways, parking, loading zones, ramps, entrances → where the van stops
● POST /v2/ortho/extract40 m tile read by Claude vision → roof tone, entrance, lot access
● POST /v2/embedding/extractSemantic search over aerial tiles → shelter leads near a patient
● GraphQL model.rooms, model.locationsRoom sizes and panoramas → crew brief, vision on the entry and lobby
● Showcase SDK sweepsEach sweep's visible neighbors → Dijkstra from the door to the device

Planner

What the profile feeds

Score
weight × hours past the deadline + 0.1 × (weight ÷ hours left) × hours waited; an unreached person costs 6 h
Pairing
Hungarian assignment, crews × people, value = drop in the score per crew-hour
Routes
Adaptive large neighborhood search: remove ~15% of stops (random, worst, cluster), reinsert cheapest, keep if better
Indoors
Dijkstra over the Matterport sweep graph, curb → door → device

Limits

What it does not know

  • Travel is straight-line, with no road, closure or flood layer.
  • No record says an elevator runs today; a resident or the building office has to say so.
  • A scan shows a unit type, never a resident's own rooms. All patients are simulated.