Business Pain Point: One Floor Plan, Ten Minutes of Manual Interpretation

A friend of mine runs a real estate SaaS mini-program serving small and medium-sized agency stores. The agents' daily workflow is: receive a floor plan photo from the landlord → visually inspect the image → manually enter structured information like "3 bedrooms, 2 living areas, north-south orientation, south-facing living room with balcony, master bedroom ~15㎡" → then write a marketing copy to post on the mini-program's listing detail page.

The pain points of this workflow are very concrete:

• Slow: For a single listing, interpreting the floor plan and writing the copy takes 8~12 minutes even for experienced agents; novices often exceed 20 minutes. A store that lists 30 new properties a day spends half a day of labor on this alone.
• Error-prone: Agents frequently make mistakes during manual entry—writing west-facing as south-facing, or two living areas as one. Wrong floor plan information leads to customer complaints during on-site viewings.
• Expensive: During peak season with high listing volume, stores either add staff or let listings go live late. Among my friend's customers, one store had a dedicated "copywriter position" with a starting salary of 6,000 RMB per month, doing only this task.

Technically, this isn't something simple OCR can solve. Floor plans contain not only text annotations but also a lot of spatial relationships: which room faces south, whether the layout separates active and quiet zones, the difference between bay windows and balconies, load-bearing wall positions. This requires a multimodal large model to reason about the image, not just recognize text.

Architecture Design: Three-Layer Structure, Deliverable by a Two-Person Team in Two Person-Months

┌─────────────────────────────────────────┐
│  小程序端（房源发布页）                    │
│  拍照/上传户型图 → 展示AI解读结果 → 人工确认│
└──────────────────┬──────────────────────┘
                   │
┌──────────────────▼──────────────────────┐
│  业务后端（Node.js / FastAPI）            │
│  · 图片预处理（压缩、方向校正、去水印增强） │
│  · 结构化输出校验（JSON Schema强约束）     │
│  · 结果缓存（同图hash命中直接返回）        │
│  · 低置信度字段标记，推给人工复核队列       │
└──────────────────┬──────────────────────┘
                   │
┌──────────────────▼──────────────────────┐
│  统一AI API网关                          │
│  · 多模态模型路由（按图片质量自动选型）     │
│  · 供应商故障自动切换 + 重试              │
│  · 统一计费、日志、限流                   │
└─────────────────────────────────────────┘

There are two core decisions:

First, don't let the mini-program connect directly to the model. All requests go through the business backend, enabling caching, validation, and a human review fallback. The probability of the same floor plan being re-uploaded isn't low (agents modify listings), so cache hits directly drive costs down.

Second, connect to models through a unified AI API gateway rather than directly integrating each vendor's SDK. This point deserves its own discussion below.

Key Implementation Steps

Step 1: Define the structured output schema (half a day)

{
  "layout": {"bedrooms": 3, "living_rooms": 1, "bathrooms": 2},
  "orientation": {"living_room": "south", "master_bedroom": "south"},
  "estimated_areas": [
    {"room": "主卧", "area_sqm": 14.5, "confidence": 0.82}
  ],
  "highlights": ["南北通透", "客厅带阳台", "动静分区"],
  "marketing_copy": "约80字卖点文案",
  "uncertain_fields": ["estimated_areas[0]"]
}

Key design: every field carries a confidence score; fields below the threshold (we set it at 0.75) go into uncertain_fields, and the frontend highlights them in yellow for the agent to verify. AI doesn't replace humans for final confirmation—it just changes the work from "filling out from scratch" to "proofreading and editing."

Step 2: Image preprocessing pipeline (1~2 days)

Photos from landlords come in all quality levels: skewed, dark, watermarked by other agencies, or screenshots with black borders. First "clean" the image with traditional image processing (orientation correction, contrast enhancement, cropping), then feed it to the model. Doing this step well noticeably improves interpretation accuracy—it's more effective than switching to a stronger model.

Step 3: Prompt and few-shot tuning (2~3 days)

Prompt essentials: explicitly state "answer only based on information visible in the image; output null for non-visible fields instead of guessing"; provide 3~5 high-quality examples covering typical layouts (one-bedroom, three-bedroom, duplex, LOFT); require distinguishing easily confused concepts like balcony/bay window and naturally lit kitchen/bathroom.

Step 4: Connect to the unified gateway + integration testing (1 day)

Configure multimodal model routing through the gateway: clear, high-resolution images go to stronger reasoning models; compressed images/screenshots go to cost-effective models.

Step 5: Grayscale rollout (1 week)

Pilot with 3 stores first, gradually reducing the human review rate from the initial 40% to below 12% before full rollout.

Why Use a Unified AI API Gateway

This small team initially connected directly to a single vendor's model API and hit two pitfalls within three months: once, vendor rate limiting caused agents to see endless spinners after uploading floor plans, and complaints went straight to my friend's company; another time, a model version upgrade slightly changed the response format, and the backend parsing crashed for an entire night.

After switching to a unified AI API gateway, the change in maintenance costs was tangible:

1. One integration codebase, multiple switchable models. Switching models only requires changing gateway configuration—no business code changes, no re-dealing with each vendor's SDK authentication and parameter differences. We estimate the adaptation effort per switch dropped from 2~3 days to half an hour.
2. Automatic failover. When a single vendor has issues, the gateway automatically routes to a backup model, invisible to the stores. For a real estate business with clear peak hours (listings published en masse in the evening), this protects the reputation.
3. Unified billing and monitoring. Multi-store account allocation, per-listing cost accounting, and abnormal call alerts—all viewable in one dashboard, without writing your own statistics logic.

For a team of two or three people, all of this would otherwise require reinventing the wheel. The gateway effectively outsources this infrastructure.

The Efficiency Math: Before and After

• Floor plan interpretation + copy per listing: ~10 minutes manual → ~1.5 minutes AI-assisted (mostly the human confirmation step)
• Store with 600 new listings per month: saves ~85 hours/month, equivalent to half a full-time employee's workload
• Error rate: manual entry occasionally produced orientation and layout errors; with the AI + review model, erroneous fields are caught and flagged in yellow at the frontend
• API cost of floor plan interpretation: spread across individual listings, only a few cents—far below the per-minute cost of human labor

---

If you're also working on a similar AI implementation project and need to access multiple models without being locked into a single vendor, you can try a unified AI API gateway: https://api.thistoken.ai/register

---

Ready to try it yourself? Sign up at https://api.thistoken.ai/register to get your API key and start building.