Zaman TauhidTechnical Lead & Software Architect
Zaman Tauhid

ResumeGitHubLinkedIn

Technical Lead & Software Architect · Dhaka, Bangladesh

Product engineering

Integrating AI into a product that already works

Start with a user task, respect the existing system, and design the data flow, validation, and operational behavior around the AI capability.

Adding AI to an existing product changes more than the interface. It introduces a new dependency into a system with established data, permissions, workflows, and user expectations.

The first architecture question is where that dependency belongs. A useful integration begins with a particular user task and defines how AI can improve it without weakening the behavior the product already promises.

Start with a task and a baseline

“Add an AI assistant” does not define an outcome. “Help a merchant find products that match a customer's description” is much more useful. It identifies the user, the information source, and a result that can be evaluated.

Establish how the task works today. What can the user accomplish? Where do they struggle? Which failures matter? A baseline may include task completion, the relevance of search results, time spent correcting generated content, or the cost of a manual operation.

Choose the measurement that reflects the product goal. A lower model latency is useful, but it does not establish that users find the right product or complete their work more successfully.

Keep the source of truth clear

An AI system may use a search index, embeddings, cached documents, or generated summaries. These are representations of product data. The application still needs a clear authority for facts such as price, inventory, permissions, and order status.

In a commerce system, retrieval can identify candidate products. Before displaying a purchasable result or applying an action, the integration should resolve the relevant current information through the product's established data pathway.

This also defines synchronization work. Decide how new records, updates, deletions, and permission changes reach the derived data. Track progress and failures so an operator can determine whether an index is current enough for its intended purpose.

Treat freshness as a design decision. A stale description and a stale inventory balance do not carry the same consequences.

Separate retrieval, reasoning, and action

These responsibilities may live in the same service, but they should remain understandable:

  • Retrieval finds relevant information.
  • Reasoning or generation uses that information to produce a result.
  • Validation checks the result against the application's constraints.
  • Execution applies an authorized action through the existing product.

For a content-generation feature, a valid result might be a structured draft. For shopping assistance, it might be a proposed cart change. Neither requires giving a model unrestricted access to the application's database.

A clean separation makes failure easier to diagnose. An irrelevant answer may come from retrieval, missing context, generation, or a mismatch between the retrieved version and current product state. Each requires a different correction.

Design the interface around uncertainty

Users need to distinguish suggestions, drafts, and committed changes. Show what will happen before a consequential action and make the resulting state explicit.

For generated content, let people inspect and edit the result. For search, offer a useful path when results are weak or missing. For long-running operations, communicate progress based on application state rather than inventing a percentage that does not represent real work.

Failure handling should respect the feature's role. If an AI description suggestion is unavailable, the existing editor can remain usable. If the feature depends on a current stock check, continuing with an unverified answer may be inappropriate.

Fallbacks need to be designed individually. There is no universal rule that every failure should silently fall back to another model or to stale data.

Plan cost and latency together

The model request is only one component of response time. Data lookup, retrieval, tool calls, validation, and delivery all contribute. Instrument the path so the team can see where time is spent.

Caching can reduce repeated work, but a cache key must reflect the context that affects correctness. A result specific to a store, user, permission set, or data version should not be reused as though it were global.

Bound expensive work. Consider request limits, timeouts, concurrency, token budgets, and cancellation. Explain these boundaries in product terms: what the user can do, when an operation may take longer, and what happens when it cannot complete.

Usage reporting should help the product owner understand which tasks consume resources and whether those tasks deliver value. Raw request counts rarely tell the whole story.

Integrate operationally, not just visually

Deployment requires the same care as the rest of the application. Document the external services, credentials, worker processes, synchronization jobs, and configuration required to run the feature.

Decide how to disable the capability if necessary. Keep ordinary application behavior available where that is possible and correct. Monitor the core task, not only whether the AI endpoint returns an HTTP success code.

For product updates, account for compatibility between the application, the integration service, and any stored data or index format. A prompt change may be easy to deploy, but it can still alter the output contract expected by the application.

Use the product boundary to guide the architecture

My own product wcbrain separates its WooCommerce integration, hosted intelligence services, and commercial platform. The plugin owns store integration and local synchronization. The hosted data plane handles discovery and AI workloads. The commercial platform manages accounts, subscriptions, and entitlements.

That separation reflects different responsibilities. It does not automatically guarantee reliability; the contracts and operational behavior between those parts still need to be designed and verified.

You can explore that product context in the wcbrain overview. The broader lesson is to make the integration's responsibilities explicit, then evaluate the feature as part of the complete user journey.

Put the ideas into practice.

Work through your development process, architecture, or next engineering challenge with me.

Explore one-to-one mentorship