DesignXEPAM Systems · 2026

AI-native design orchestrator

Design as an agentic pipeline

not a prompt

My role

Design lead. I led DesignX from an npm package experiment to enterprise delivery: the pipeline and its gates, a team of 15+ contributors and every screen of the product, designed and built myself.

Used in enterprise delivery by EPAM designers working for

In
a regulated fintech brief
Agents
11
Phases
6
Out
a design system of record

01  ·  The gap

AI made design faster and impossible to defend.

By 2026 generative tools could produce a screen in seconds. The market split into two camps, and neither one produced the thing a senior team actually needs.

Camp one

One-shot generators

v0, Lovable, Figma AI. A screen in seconds, with no research lineage and no rationale, and they will invent facts to fill the gaps.

Camp two

Copilots

Genuinely useful for one artefact at a time. But what to make, in what order, gated by what: all of that still lives in your head.

Neither

A system of record

Nothing produced a durable chain: the artefact, the insight beneath it and the evidence beneath that. Output without lineage.

The cost

Undefendable work

A team cannot ship what it cannot defend in a review. Speed that produces unciteable work is not speed.

02  ·  The bet

What if design ran like a production line?

Not a prompt you fire and hope. A governed pipeline: distinct phases, specialised agents, hard approval gates and an evidence contract where nothing ships unsourced.

What was on offer  ·  one hop
briefa screenno lineageno way back
What we built  ·  a governed line
brief00onboard01discover02ideate03design04deliver05measurea design systemof recordre-runnableHITL gates, each held until a human clears it

03  ·  My role

I owned the line, from strategy to the last pixel.

Depth of ownership across six areas

  1. Co-led1
    • Orchestration and control modelRun modes, autonomy levels, model routing and context budgets.
  2. Led3
    • Product strategy and pipeline shapeThe six phases, the nine gates and the reframe from prompt to production line.
    • Agent and skill definitions11 agents across 13 categories, plus the protocol every agent inherits.
    • The quality barA 482-line design authority that grades every visual output and the eval harness that enforces it.
  3. Sole2
    • Token optimisationAppropriate model used per task with BYOK, Ollama and LM Studio support built in.
    • The entire product UXEvery screen in this case study. Design and build, start to finish.

Four areas led  ·  two owned outright

Sole means exactly that: no other designer touched the product surface

Leading the team

Most of the 15+ people I led did not report to me.

01  ·  The team

15+ contributors, their own projects

Contributors from across EPAM, each working on their own projects alongside DesignX. A couple reported to me; most did not.

02  ·  The rhythm

Alternate-day stand-ups

Weekly critiques and async channels kept the work moving between them.

03  ·  Stakeholders

Milestone demos

Progress shown to stakeholders at each milestone and at the quarterly design all-hands.

04  ·  Engineering

One pipeline, two disciplines

Worked with engineering teams to route contributions into EPAM’s agentic delivery pipelines.

I ran the team the way the pipeline runs: work comes in, gets reviewed against a bar and only then merges.

04  ·  The pipeline

Six phases. Multiple gates. Nothing skips.

A phase cannot advance until its gate clears and a human decides when it does.

  1. Phase 00  ·  Gate DSG00

    Onboard

    Three calm questions and a knowledge-base upload. Name it, pick a run mode, drop in what you already know. Everything the pipeline needs, before any generation happens.

  2. Phase 01  ·  Gate D101

    Discover

    Agents read the uploaded corpus and produce insights with evidence IDs and confidence levels attached. Nothing is asserted without a citation back to a source in the knowledge base.

  3. Phase 02  ·  Gate D202

    Ideate

    Concepts are generated against the insights, then scored against a requirements checklist. Unmet requirements and blockers are on the table before the gate opens.

  4. Phase 03  ·  Gate VDG03

    Design

    A design system of record is produced first: tokens, components, states. Screens are then specified against it, and a Design Director pass reviews visual quality before anything is called done.

  5. Phase 04  ·  Gate D304

    Deliver

    The Builder turns the design system and screen specs into real, running code, opened in a live preview beside the conversation. Spec and product stay in the same window.

  6. Phase 05  ·  Gates R1 to R405

    Measure

    A designer-QA rubric scores the output across eight dimensions and names every gap it finds so each one can be resolved to a production spec.

05  ·  Progressive disclosure

Hiding a studio behind three questions.

The hardest design problem was not the machinery. It was making the machinery feel like nothing. Every one of the agents and skills stays invisible until the moment it is needed.

  1. New project · GreenBasket (demo only)
    DesignX new-project screen: the assistant asks what is being built, with a name field reading GreenBasket and a description field describing a London-based online marketplace connecting urban consumers with sustainable local producers.

    GreenBasket  ·  Test project

    A fictional London grocery marketplace, set up to demo the pipeline end to end for the purpose of this case study.

DesignX started as a directory you worked in from VS Code, which suited developers more than designers. I pushed for an app designers would actually open, led it and onboarded them myself: demos for most and workshops for those who preferred to stay in VS Code.

06  ·  Control

Autonomy is a dial, not a switch.

Trusting a pipeline is not all-or-nothing. Four run modes move the same pipeline along a single axis: how often it stops and asks you. The gates never disappear; they only change who opens them.

Run mode

Gate behaviour

Phase boundaries pause

When you reach for it: The default.

Barred = pauses for you  ·  Hollow = clears itself

Control is only half of trust. The other half is proof.

07  ·  Traceability

Every pixel can name its source.

GreenBasket is a fictional London grocery marketplace, set up to test the pipeline. One of its screens, traced backwards to the survey row it came from.

  1. Stage 1 of 4  ·  Evidence  ·  E22 · E36

    Two rows in a survey

    Consumer survey data and a user transcript. GreenPass converts at 26.3% against a 35% target. Tom, a subscriber, is unsure whether the membership saves him money on this specific order.

    Artefact · E22 · E36
    A DesignX evidence artefact: a table of metrics, each row carrying its own evidence ID such as E05, E23 and E32.
    Stage
    Evidence
    Phase
    01 Discover
    IDs
    E22, E36
  2. Stage 2 of 4  ·  Insight  ·  INS-08

    The value is invisible at the decision

    The subscription offers genuine economic value, but it is never shown at the moment it matters: checkout, when the delivery fee appears. Confidence HIGH, sourced to E22 and E36.

    Artefact · INS-08
    The INS-08 insight artefact, headed “GreenPass Value Proposition Is Invisible at the Moment of Decision”, showing its evidence IDs, confidence level and design implication.
    Stage
    Insight
    Confidence
    HIGH
    Sourced to
    E22, E36
  3. Stage 3 of 4  ·  Design implication  ·  D3

    No dark patterns; just honest arithmetic

    The artefact does not stop at the finding. It writes the rule: at checkout, show non-subscribers what they would have saved on this order and nothing more manipulative than that.

    Artefact · D3
    A DesignX artefact panel showing the “SOURCED” provenance contract: will adapt, not recreate.
    Stage
    Implication
    Derived from
    INS-08
    Contract
    Sourced, not recreated
  4. Stage 4 of 4  ·  The pixel  ·  Built

    Arithmetic, on the checkout screen

    The order summary states subtotal, delivery and total with nothing hidden, then reports the basket’s Impact Score. The line traces back, unbroken, to a row in a spreadsheet.

    Artefact · Built
    The GreenBasket checkout order summary as built: subtotal £21.20, estimated delivery £4.80, total £26.00, a place-order button and a basket impact reading of Deep Green 82.
    Stage
    Built
    Traces to
    E22 via INS-08
    Screen
    Checkout summary

Real output from the GreenBasket run, not a reconstruction

08  ·  The eval

A tool that grades its own work at five out of five is not an eval, it is marketing.

Generation is easy to fake and hard to trust. So the pipeline scores its own output against a designer-QA rubric across eight dimensions and reports the low scores as loudly as the high ones. The GreenBasket run came back at out of 5 with 16 named gaps, every one resolved to a production spec before delivery.

Designer-QA rubric · eight dimensionsMean 4.1 / 5

Scored 5  ·  Two dimensions

  • Content & copy
  • Token discipline

Scored 4  ·  Five dimensions

  • Visual hierarchy
  • Accessibility
  • Responsiveness
  • Component structure
  • Performance

Scored 3  ·  One dimension

  • Interaction detail
Nothing scored 1 or 2. The single 3 is where the next run starts.

The lowest score is the point

Interaction detail scored 3. That is the number I care about most.

The rubric exists to produce a worklist and a 3 is where the next run starts. Each gap arrives as a named, specified fix: not “improve the interactions”, but the component, the state and the intended behaviour.

09  ·  The record

The deliverable is not the screen. It is the system that produces screens.

A one-shot tool hands you a picture. DesignX hands you the argument for every pixel in it and the system that will produce the next hundred consistently.

Stage 01

Evidence

Every source in the knowledge base, indexed and addressable by ID.

Stage 02

Insights

Findings that cite evidence IDs and carry an explicit confidence level.

Stage 03

Design system

Tokens, components and states, produced before any screen is drawn.

Stage 04

Production specs

Screens specified against the system, each traceable to the insight beneath it.

Artefact · design system
A DesignX design-system artefact listing resolved tokens, component definitions and their states.
Phase
03 Design
Gate
VDG
Produced
Before any screen
The design system is generated as an artefact in its own right, with resolved token references, not implied by the screens.
Artefact · design principles
A DesignX artefact listing the design principles derived for the project, each tied back to research findings.
Phase
03 Design
Derived from
Research, not a template
Screens argued against
All
Principles come out of the research, not out of a template, and the screens are then argued against them.

10  ·  The output

From brief to a working app.

Opportunity
Grocery delivery as convenient as your usual shop and as meaningful as your local market.
Problem
Trust and how easily a confusing label or a silent substitution erodes it.
Answer
The three decisions that follow are the run’s answers.
Screens
Untouched: a working front end built from the generated design system, with real copy and real product logic.
GB · homepage
The full GreenBasket homepage, top to bottom: the hero, headed “As convenient as your usual shop. As meaningful as your local market,” with body copy citing 127 verified producers across four London boroughs and a “See what's in season” button; four trust cards; this week's harvest as a grid of six products, each with an Impact Score; a producer story on Marsh Lane Growers; shop by category; a GreenPass membership banner; and the footer.
Phase
04 Deliver
Gate
D3
Built from
The generated design system
Positioning, not filler. The pipeline wrote a claim it could substantiate: 127 verified producers across four named boroughs, because the knowledge base gave it the numbers to be specific with.
GB · order confirmation
The GreenBasket order confirmation screen, showing the order reference, delivery window and itemised basket.
The run ends where a real one would: a confirmation screen, not a hero shot.

Three decisions worth pointing at

Every claim has a “Learn more”

Four trust claims sit above the fold, and not one of them is a bare assertion. Each opens its own methodology. “No dark patterns. Ever.” is an unusual thing for a generated store to promise, and it traces directly to the INS-08 implication.

GB · trust claims
A row of four trust cards on the GreenBasket homepage: 127 verified producers, Impact Score, Allergen data and “No dark patterns. Ever.”, each with explanatory copy and a Learn more link.

Substitution is a preference, not a surprise

Every basket line carries its own “if unavailable” choice: substitute, contact me first, or remove. Fresh-grocery delivery lives or dies on this interaction, and the pipeline surfaced it per item rather than burying one global setting in an account page.

GB · basket
The GreenBasket checkout basket: two line items, each with an “If unavailable” row offering Substitute with similar, Contact me first, or Remove if unavailable.

Allergens on the card, not the detail page

Products are checked against the 14 major EU allergens, and the warning renders on the card itself. The Impact Score badge sits in the same corner on every tile: the kind of consistency that comes from specifying a component before drawing screens.

GB · product grid
The GreenBasket product grid: six product cards each showing an Impact Score badge, producer name, weight, price and, where relevant, an allergen warning such as “Contains: Gluten (wheat), Sesame”.
GB · checkout commitments
A block of four checkout reassurances: No hidden fees, 14-day cancellation right, Payment used only for this order and No dark patterns. Ever.
At checkout, the same four commitments reappear, including the UK distance-selling cancellation right, stated in plain language.

11  ·  Proof

npm downloads, May to September 2026

Proof it wasn’t a prototype.

I pushed to publish DesignX on npm early, to test it with a community. The answer was clear: a real opportunity.

Check the npm package (opens in a new tab)

Shipped

71 releases in 22 days

Runs locally

Nothing sent to an outside server

Copilot, Ollama, LM Studio or a team’s own key.

Client work

Now

Open source

No longer maintained on npm.

The lasting idea isn’t the agents. It’s that when design becomes a pipeline, the deliverable stops being a screen.

It becomes the system that produces screens: every decision traced, defended and re-run.