From Lab Experiment to Mobile Feature: Building AI Products That Actually Work

The gap between impressive AI research and reliable consumer products is vast. Bhavuk Jain, speaking at InfoQ, breaks down the systematic process Google uses to transform raw foundational models into shipping features—a framework that applies equally to generative tools like AI wallpapers and intuitive features like Circle to Search.

The Four Pillars of AI Product Development#

Modern AI doesn't start from scratch. Instead, engineers inherit a massive, unpredictable foundational model and refine it through four essential stages:

Post-training teaches the base model to align with human preferences, style, and safety constraints—much like coaching a brilliant but untamed student. Two main approaches compete here: RLHF (reinforcement learning from human feedback) offers flexibility but complexity, while Direct Preference Optimization (DPO) trades some expressiveness for simplicity and scalability.

Fine-tuning transforms a generalist model into a specialist. Supervised fine-tuning on domain-specific data remains standard, but newer Parameter-Efficient Fine-Tuning (PEFT) methods like LoRA freeze most original weights while training lightweight specialized modules. This dramatically cuts training cost and speed—though sometimes the real answer isn't more training at all, but grounding the model in fresh external knowledge.

Retrieval and grounding connect the model to external sources of truth—structured databases, web results, or on-device context—preventing hallucinations and keeping outputs factual and current. This layer has become essential as the field discovered that even fine-tuned models can confidently invent facts.

Inference, guardrails, and deployment wrap everything in multilayered safety. Data guardrails filter toxic or sensitive content upstream; model guardrails reshape behavior through safety-tuned weights and classifier heads; runtime guardrails inspect inputs and outputs for PII and jailbreak attempts in real time; and governance layers add human review and user controls. Only then can infrastructure serve the model reliably across millions of requests.

AI Wallpapers: Engineering Creativity at Scale#

Building a feature that lets users generate personalized wallpapers on their phones exposed two competing design philosophies: maximum freedom versus guided simplicity.

The blank-canvas approach—giving users an open text box to describe anything—sounds powerful in theory. In practice, it shifts the burden onto consumers to become prompt engineers and creates a non-deterministic product impossible to guardrail reliably. Jain's team faced user frustration in existing public products and made a counterintuitive bet: rather than maximum complexity, they chose . Users select from beautiful pre-made concepts and refine them, keeping quality predictable while preserving creative control. This isn't a limitation—it's an engineering choice that makes reliable guardrails possible.