7‑Step Architecture Framework for AI Product Managers to Build Scalable Solutions

As large‑model technology matures, enterprise AI applications face deep‑water challenges: many AI products remain cheap API‑wrappers that fail to integrate with core business flows, resulting in poor ROI. The author recounts the internal project "SmartPhoto Cross‑Border E‑commerce AI Shooting System" (2023) to illustrate how a senior AI product manager can move from prototype to industrial‑grade deployment.

1. Business Baseline and Pain‑Point Definition

Cost baseline: Sample logistics, studio scheduling, on‑site setup, photography and post‑processing amount to ¥65 per SKU main image.

Cycle baseline: End‑to‑end average time 3–5 days, peaking at 9 days during peak sales.

Capacity & loss: Physical bottleneck limits output to ~10,000 images/month, costing >¥600,000, while slow rollout wastes high‑frequency testing opportunities.

The core mission was to build an automated system that reduces per‑image cost to ¥1.0 and delivery time to minute‑level , while preserving or exceeding the visual conversion baseline.

2. Demand Decomposition & Stakeholder Analysis (Step 1)

Using a "5 Why" analysis, the author uncovered that the real deficiency is the lack of a low‑cost method to reconstruct realistic physical lighting. Generating images without accurate light and shadow would destroy conversion rates, even if thousands of images are produced per minute.

3. Technical Selection & Feasibility Assessment (Step 2)

External SaaS/API options (e.g., Photoroom, Midjourney API) were tested and found to have severe feature‑generalisation issues for e‑commerce categories, yielding a usable rate of under 30% . Their token‑based pricing also caused linear cost explosion at the required 100k+ monthly calls and raised data‑compliance concerns.

Open‑source Stable Diffusion WebUI offered low compute cost but required operators to master complex parameters (CFG Scale, Denoising Strength), leading to near‑zero internal adoption.

Conclusion: the MVP must strip away B‑class high‑precision demands and focus on A‑class high‑throughput pipelines.

4. MVP Boundary Definition & Category Control (Step 3)

An In‑Scope/Out‑Scope matrix was created:

In‑Scope (Do): High‑conversion home and pet products with uniform visual features.

Out‑Scope (Postpone): Clothing (complex human pose, fabric folds) and 3C electronics (metallic reflections) due to high risk of feature dilution.

Functional scope was limited to automated background removal, a curated high‑frequency scene library, and an asynchronous high‑concurrency generation queue. Custom prompt editors, local mask re‑draw, and AI‑driven model‑swap were explicitly rejected.

5. Product Architecture & Interaction Abstraction (Step 4)

Because large models are probabilistic, the product must present a deterministic front‑end. The system hides the stochastic backend behind a single "click‑to‑generate" button. When a user selects a thumbnail (e.g., "Nordic Morning Light"), the backend automatically assembles a complex prompt stack, injects a negative‑prompt library, applies category‑specific LoRA weights, and sets appropriate diffusion steps, delivering an industrial‑grade "no‑brain" experience.

6. Metric Funnel & Risk‑Control System (Step 5)

A three‑layer data funnel (L1‑L3) was instrumented to monitor model health and ROI, with a north‑star usable‑rate target of 72% . The funnel defines:

L1: Raw generation success.

L2: Business‑level usable images (≥72% required).

L3: Cost‑per‑image and time metrics.

Gray‑release follows a "3‑3‑1" rule:

Gate 0 (Internal test): 20 seed users to probe extreme tail data.

Gate 1 (Differentiated test): 80 users from high‑value teams; if L2 drops below 72%, immediate physical melt‑down and data‑augmentation are triggered.

Gate 2 (Full rollout): All lines after L2 stays ≥72% for five consecutive workdays, accompanied by SOP release.

A hard stop line: if after 10 weeks the core category usable rate stays below 60%, the project is terminated and the workflow reverts to traditional outsourcing.

7. Data‑Driven Gray Release Strategy (Step 6)

The same gate framework enforces incremental exposure, ensuring that each stage validates both technical stability and commercial ROI before scaling.

8. Continuous Architecture Evolution & Artifact Archiving (Step 7)

Post‑launch, a feedback‑to‑layer routing map directs user‑reported issues to the appropriate stack:

Presentation layer: Bulk‑upload requests trigger engineering work on asynchronous batch modules.

Probability layer: Light‑edge artifacts lead to inference‑layer upgrades, such as integrating ControlNet depth constraints.

Feature layer: New material failures feed the training pipeline, prompting LoRA fine‑tuning once a bad‑case threshold is reached.

Results: per‑image compute cost dropped to ¥0.5 , delivery time compressed from 3 days to 1 minute**, and business line penetration exceeded 65%, freeing 30 designers from manual shooting.

Appendix: Core Deliverables of a Senior AI PM

Stakeholder & Pain‑Point Analysis Matrix.

Technical Selection & ROI Accounting Model.

MVP Category & Feature Control (In/Out) Matrix.

Data Funnel & Monitoring Definition Document.

Gray‑Release & Melt‑Down Plan.

Architecture Evolution Roadmap linking front‑end complaints to engineering, inference, and training improvements.