Real Estate AI Agent Training Environment: OpenEnv-Based Hybrid Residential and Commercial Property Workflow Simulator

This article introduces the open-source project openenv-hackathon-team-ai-alchemist, an OpenEnv-style training environment for the real estate sector that simulates the real workflows of residential sales and commercial property leasing, providing AI agents with a testing and training platform close to actual business scenarios. The environment features strong-typed data models, fine-grained reward functions, multi-difficulty task levels, and real-time visualization tools, supporting AI agents in learning key business operations such as lead screening and property matching.

Traditional AI benchmark tests mostly focus on games or abstract tasks, lacking simulation of real business workflows. Real estate sales teams need to handle a large number of customer inquiries, complete complex decisions such as lead screening, follow-up visits, and property matching, and often make reasonable judgments when information is incomplete. This project aims to fill this gap by creating a non-game training environment for AI agents to learn real business operations: rewards are based on pipeline quality improvement and expected revenue contribution, while penalties target inappropriate recommendations, premature stage advancement, etc., ensuring strategies translate into actual business value.

The core class of the environment is RealEstatePipelineEnv, which uses Python + Pydantic to implement strong-typed data models, including three models: Observation (decision context: task ID, step count, pending queue, etc.), Action (nine operations: classify opportunities, set priorities, request missing information, recommend properties, etc.), and Reward. The nine actions cover the entire lifecycle of real estate sales, reflecting the multi-step, multi-round interactive decision-making mode of real teams.

The reward function provides fine-grained learning signals: positive rewards include correct customer classification, reasonable priority setting, successful property matching, etc.; penalties target invalid actions, inappropriate recommendations, unrealistic rental terms, etc. Evaluation is done by a deterministic scorer, returning a normalized score between 0 and 1, with three difficulty tasks (residential buyer qualification verification / missing information follow-up / commercial rental strategy, corresponding to easy/medium/hard).

The project includes a CRM-style traffic simulator that demonstrates the complete residential buyer service process; supports FastAPI endpoints for submitting custom leads, streaming API for pushing NDJSON real-time data, and a browser dashboard for visualizing decision processes; the command-line tool scripts/stream_live_traffic.py can display multi-lead processing flows in real time, facilitating debugging and demonstration.

Built-in training data generation pipeline supports step-level (decision point context + action target) and task-level (full trajectory result) data export; generated records include tags such as query text, lead profile, inventory snapshot, etc., which can be directly used for supervised fine-tuning; baseline model training scripts are provided, supporting residential/commercial classification and next-action prediction model training.

Applicable to real estate tech companies training customer service automation agents, CRM vendors verifying AI functions, research institutions exploring vertical applications of LLMs, and educational institutions designing practical AI courses. Its design patterns (strong-typed models, fine-grained rewards, multi-difficulty tasks, visualization tools) can be migrated to other vertical fields, providing methodological references for industry AI applications.

This project represents the evolution direction of AI agent training environments from general games to industry real simulations. By accurately modeling real estate processes, it provides AI agents with training standards close to reality. Its modular architecture and complete toolchain reduce the development threshold of industry-specific training environments, and future vertical field-specific training environments will become key infrastructure for AI implementation.