Skip to main content
AI · LLM · AgentsAvailable now · 1 slot open

AI product UX designer for products people actually use.

Most AI products feel like a dashboard someone is afraid of, or a chat window pretending the rest of the product doesn't exist. I design AI tools — chatbots, LLM products, agent interfaces — that hide the right complexity, surface the right uncertainty, and don't lose users in the first 90 seconds.

Hire me if you're building an AI chatbot, an LLM-powered tool, an agent flow, or any AI interface where the user has to trust a non-deterministic system. I bring principled patterns from shipped product work in crypto and ecommerce — categories whose UX problems are structurally the same as AI's.

See the closest case study
Why AI product UX is hard

A new category with no settled patterns.

The category is too new for cargo-culted patterns

Crypto inherited fintech's patterns. SaaS inherited desktop's. AI products arrived with nothing — three of the dominant UIs (ChatGPT, Claude, Perplexity) were designed in the same 18 months and copy each other. That's not a settled grammar; that's a draft. Building on those defaults uncritically locks you into interactions that nobody has stress-tested at your category yet.

The interface has to hide complexity without dumbing the product down

The user shouldn't see system prompts, model selection, temperature sliders, or retrieval breakdowns on the default surface — but the power user needs all of it reachable. Designing the boundary is the work. Most AI products either expose everything (overwhelming) or hide everything (infantilising). The right answer is layered — and the layers have to match how your specific users think, not how the model thinks.

Trust + uncertainty signals matter more than in non-AI products

In a non-AI product, the system is deterministic — the user trusts the interface because the interface is reliable. In an AI product, the system is probabilistic — the user has to calibrate trust per-answer, sometimes per-sentence. That's a UX problem the field hasn't solved yet. Citation chips, confidence badges, and “draft” framing are the early grammar — they need to be designed at the system level, not bolted on per-feature.

Streaming and non-deterministic outputs break standard loading patterns

Every design system has a “loading → result” pattern. AI products have “loading → partial result → still loading → final result, sometimes wrong, sometimes retried.” The standard spinner is a lie. Streaming text needs its own visual treatment, cancel needs to be reachable mid-stream, and the user needs a read on whether what they're seeing is a partial answer or a stalled one. None of that comes out of the box.

Where the hard work lives

What I focus on in AI product UX.

  • Conversational interfaces beyond the bubble

    Chat is a primitive, not a layout. I design conversational UI that earns its bubbles — inline editors for long prompts, structured replies with checklists and tables, slash-menu vocabularies, and quiet history so the screen doesn't turn into a wall of text.

  • Confidence + uncertainty signals

    The hardest UX problem in AI products: when is the model sure, when is it guessing, and how do we tell the user without making the interface feel apologetic. I design the visual grammar — citation chips, confidence badges, "draft" framing — that lets users calibrate trust without reading the docs.

  • Multi-step agent flows with editable handoffs

    Agents that plan, then act, then verify need handoff points the user can intervene at. I design the plan-preview, the per-step approval, the resumable failure state — so an agent run feels less like a black box and more like a colleague checking in.

  • Onboarding for users who don't trust AI yet

    Most AI products onboard like they're onboarding researchers. Real users arrive sceptical. I design the first 90 seconds — what to show, what to defer, where to put the "this is what just happened" explanation — so non-technical users get to a value moment before the trust budget runs out.

How it works

Four steps to shipped.

  1. Discovery call

    30-min free call. We figure out if we're a fit and what the engagement looks like.

  2. Fixed-scope proposal

    Within 48h, you get a written scope with timeline and price. No hourly billing.

  3. Kickoff + 6-phase build

    Research → Wireframes → Visual → Test → Handoff. Weekly Loom, async Figma.

  4. Handoff + 2 weeks Q&A

    Dev specs, animation specs, 2 weeks in your engineering Slack.

See the full 6-phase process
The closest case study

Enrichplay — the same problem shape as AI products.

The design problem I solve in crypto is structurally the same as the design problem in AI — dense data, complex state, users at very different literacy levels, and a default vocabulary that assumes prior knowledge. The principles transfer.

Enrichplay was designed for the crypto-curious, not the crypto-native. Plain language at the boundary, never inline. Explain WHY before HOW. Defer power-user controls until asked. Surface the one thing that matters next to every number. The same instinct is what makes an AI product feel approachable instead of intimidating.

Read the Enrichplay case study
  • Dense data, surfaced without overwhelm
  • Default vocabulary tuned for newcomers
  • Power-user controls deferred, not hidden
  • One responsive system across mobile + web
  • Tabular figures so numbers don't reflow
  • Tone: confident, not condescending
Pricing

How to work together on AI products.

Three engagements — a Sprint for a single flow (chat, agent handoff, onboarding), a Full Project for end-to-end AI product UX, or a Retainer for ongoing iteration as the model and product co-evolve.

  • D
    1–2 weeks

    Design Sprint

    Quick win, ship in 2 weeks.

    $450

    A focused engagement to fix a specific flow, audit an existing product, or design a single feature end-to-end.

    • 1–2 weeks of focused design time
    • 1 specific flow OR a heuristic audit
    • Hi-fi screens for desktop + mobile
    • Interactive Figma prototype
    • Async Loom walkthrough at delivery
    • 1 round of revisions
    Best forFounders who need a specific thing done well, fast — landing page, onboarding flow, settings redesign, or an audit before raising.
  • Most chosen
    F
    4–8 weeks

    Full Project

    End-to-end, research to handoff.

    $800–$1,200

    The full engagement — user research, wireframes, hi-fi screens, prototype, usability testing, and developer handoff.

    • Discovery interviews + research synthesis
    • Wireframes for every flow
    • Hi-fi design (iOS + Android OR desktop + mobile)
    • Design system + reusable components
    • 5-person moderated usability test
    • Interactive prototype + walkthrough
    • Dev handoff specs + 2-week Q&A office hours
    • Up to 3 rounds of revisions
    Best forFounders shipping v1 or doing a serious v2 redesign. The work that needs to feel right, not just look right.
  • M
    Rolling, 2-month min

    Monthly Retainer

    Ongoing design partner.

    $500/mo

    Reserved hours every month for ongoing design work — new features, iteration, system maintenance, async reviews.

    • 40 reserved design hours / month
    • Weekly 30-min sync
    • Async Figma + Slack throughout
    • Design-system governance + extensions
    • Roll-over up to 8 hours / month
    • Priority response (< 24h on weekdays)
    Best forFunded startups (seed to Series B) with a roadmap that needs a consistent design hand — and a founder who values speed over RFP processes.

Full pricing breakdown, side-by-side comparison, and add-ons live on the pricing page.

Recent projects.

See all case studies
10+
Products Shipped
4+ Years
UI/UX Experience
5
Industries Designed For
4+
Countries Served
Services

What else I design.

AI-product-specific questions.

  • Honest answer: AI is one of the industries listed on my homepage stats — but my directly-named, public case studies are crypto (Enrichplay), ecommerce (Painted Juttay), and healthcare (Take Therapy UX audit). The principled approach transfers cleanly to AI products because the underlying problem shape is the same — dense state, intimidating defaults, users at very different literacy levels — but I won't claim shipped AI client work I can show off if I haven't shipped it. I'll be upfront on the discovery call about which patterns I'm bringing from analogous categories vs. inventing in your domain.

  • I start by separating "chat" from "conversation." Chat is a layout — bubbles, an input, scrolling history. Conversation is a structure — turns, intent, repair, hand-offs. Most AI products copy the chat layout from ChatGPT and skip the conversation design underneath. I do it the other way around: map the conversation shape first (what intents, what edits, what dead-ends), then choose the layout that fits — sometimes that's bubbles, often it's a document with an AI rail, an editor with inline suggestions, or a structured form the AI fills in. Layout is a consequence, not a default.

  • Three layers, and I design all three. First, the visual grammar — confidence chips, citation pills, "draft" treatment for low-certainty content. Second, the framing — "here's my best guess" vs. "here's the answer" reads as a completely different product. Third, the recovery affordance — when the model is wrong, can the user steer it back in one move or do they have to start over? Most AI products do layer one, half of layer two, and skip layer three. Skipping layer three is what burns trust the fastest.

  • Yes — and I prefer engagements where I can. I don't train models, I don't write evals, and I won't pretend to. But I'll learn what your model does well and badly, design the UI to flatter its strengths and surface its weaknesses honestly, and loop your AI team in at wireframe stage so we don't design a flow your model can't actually support. The bad version of this collaboration is "design ships a Figma file, eng can't build it." I avoid that by sketching with eng in the room from week one.

  • Streaming breaks the standard "loading spinner → result" pattern most design systems are built around. I design for the in-between state — partial text that the user can read as it arrives, cancel affordances that don't feel destructive, "still thinking" states that aren't fake spinners. For multi-step agent flows, the harder question is where to put the handoffs: per-step approval is safe but slow, end-of-run review is fast but unrecoverable. I design tiered handoffs — light approvals for reversible steps, explicit confirms for irreversible ones — borrowing the same Nielsen-5 logic I used for destructive actions in the Take Therapy audit.

  • Yes — through analogy. Crypto products (Enrichplay) are the closest structural match to AI products on the user side: dense data, intimidating defaults, a vocabulary that assumes prior knowledge, and a user base split between people who know what they're doing and people who arrived two months ago. The Enrichplay design system was built for the crypto-curious, not the crypto-native — plain language at the boundary, never inline; explain WHY before HOW; defer power-user controls. The exact same instinct applies to onboarding non-technical users into an AI tool.

Related

Also explore.

Available now · 1 slot open

Got an AI thing to ship?

Or email me
Based in
Karachi, PakistanUTC+5