Skip to content
Md Shariful Islam

AI integration

AI inside real products — not beside them.

I add AI to software where it creates measurable value for users, integrated with the application's existing data, workflows and security model. Two shipped projects illustrate the two main approaches: cloud-based generative AI and on-device machine learning.

Implemented examples

Cloud generative AI

Guroos — OpenAI-assisted author profiles and content

Within a headless Next.js and WordPress marketplace, the OpenAI API assists with generating author-profile and supporting content, so experts are not starting from a blank page.

Runs
Cloud API (OpenAI)
Model type
Large language model
Integrated with
Next.js, headless WordPress, Node.js
Best for
Language generation inside existing workflows
Read the case study

On-device machine learning

PlantDoc — offline plant disease classification

An Android app that runs a MobileNetV2 image classifier through TensorFlow Lite directly on the phone, returning a likely disease and confidence score without a network connection.

Runs
On the device, offline
Model type
MobileNetV2 image classifier (TFLite)
Integrated with
Kotlin, Jetpack Compose
Best for
Fast, private inference without connectivity
Read the case study

Choosing an approach

Cloud generative AI vs on-device ML

They solve different problems. Picking the wrong one leads to slow, expensive or privacy-unfriendly features.
Comparison of cloud generative AI and on-device machine learning
AspectCloud generative AIOn-device ML
Where it runsProvider's servers, via APIOn the user's phone
ConnectivityRequires a network connectionWorks offline
StrengthsBroad language understanding and generationLow latency, privacy, no per-request cost
ConstraintsLatency, usage cost, data leaving the appModel size and hardware limits
Fits whenGenerating or transforming text in a workflowFast classification at the point of need
ExampleGuroos — author-profile contentPlantDoc — leaf disease classification

Engineering practice

How I approach AI integration

The principles I apply when adding models to production software.

  1. Model and API integration

    AI calls live behind the application's own server-side boundary, with credentials kept out of the client and usage tied to real product actions.

  2. Structured outputs and validation

    Model output is treated as untrusted input: requested in a defined shape and validated against a schema before it reaches storage or the UI.

  3. Error handling and reliability

    Timeouts, retries and clear fallbacks mean a slow or failed model response degrades a feature — it does not break the workflow around it.

  4. Fit with existing workflows

    AI is added where users already work — inside a CMS, an onboarding flow or a mobile screen — rather than as a separate tool to learn.

  5. Security and user data

    Send the minimum data needed, avoid exposing personal information unnecessarily, and prefer on-device inference when data should not leave the user's device.

  6. Practical use cases first

    Start from the problem — a cold-start profile, an offline diagnosis — and choose cloud generation, on-device ML or no AI at all accordingly.

Before we build

Questions I ask before adding AI

  1. What decision or task does AI make faster, better or possible — and how will we know?
  2. Does the data need to leave the user's device or your infrastructure at all?
  3. What happens when the model is slow, unavailable or wrong?
  4. Where will humans review, edit or override the output?
  5. Is a deterministic rule or a search index a better fit than a model?

Let's talk

Want to add AI to a product that already works?

Discuss a freelance or consulting project

A platform to build, a legacy system to modernise, an integration that has to work, or AI to add to an existing product.

Start a project inquiry

Connect about a senior engineering role

Remote senior full-stack positions where ownership, product judgement and breadth across the stack matter.

Talk about a role

Prefer email? kaitangroup@gmail.com