Zia ArsalanHome›AI Product Development
AI Development
AI features that survive contact with users
Most AI projects demo well and disappoint in production, because the hard part is not the model — it is the data around it, the cost per request, and what happens when the answer is wrong. I build AI into products that people use every day, and I will tell you when you do not need AI at all.
AI products I have shipped
What I build
AI features inside an existing product
The common request: you have a working product and want one thing in it to be intelligent. Summarising, classifying, extracting, matching, drafting. Scoped narrowly so it ships.
Support and sales chat
A chat assistant grounded in your own documents and data, that answers accurately and escalates rather than inventing an answer.
Document and data extraction
Turning invoices, CVs, forms and PDFs into structured data your systems can actually use.
Workflow automation
Routing, triage and enrichment steps that used to need a person reading every item.
AI in your CRM
GoHighLevel and similar platforms, wired to models so automations can make decisions rather than just fire on triggers.
A second opinion
If you are unsure whether AI is the right tool for a problem, I will tell you when a rule, a search index or a form would do the job better and cheaper.
How I approach it
The unglamorous parts decide whether it works
Data quality, retrieval and evaluation determine whether an AI feature is useful. Model choice matters far less than most demos suggest.
Wrong answers cost more than no answers
A feature that is confidently wrong damages trust faster than a missing feature. Anything shipped needs a way to say "I do not know" and a path to a human.
Cost is a design decision
Token spend scales with usage. Caching, model routing and keeping prompts tight are part of the build, not an optimisation for later.
Common questions
Which models do you build on?+
Whichever fits the job and the budget — commercial APIs for most work, smaller or self-hosted models where cost, latency or data residency demands it. The model is the easiest part to change later, so it is not the first decision.
Will my data be used to train someone else's model?+
Not if it is set up properly. Business API tiers do not train on your data by default, and where that is not acceptable at all, the work runs against models you host. This gets decided at the start, not after launch.
How long does an AI feature take to build?+
A narrow, well-defined feature inside an existing product is typically 2 to 4 weeks. A standalone AI product is a normal product build with an AI component in it, so it runs on normal product timelines.
How do you stop it making things up?+
By grounding answers in your own content rather than the model's memory, constraining what it is allowed to answer, and testing against a fixed set of real questions before launch. Hallucination is reduced by architecture, not by asking the model nicely.
Is AI actually the right answer for my problem?+
Often it is not. If a lookup table, a search index or a better form solves it, that will be cheaper to run and easier to maintain. I would rather say so early than build something you regret paying to keep running.
Describe the problem, not the AI
Tell me what a person on your team spends too long doing, or what your customers keep asking for. I will come back with whether AI solves it, what it would take, and what it would cost to run.