The Adaptive Experience Platform The Adaptive Experience Platform
Lyra Agentic is an adaptive user interface for NICE's Customer Experience Analytics (CEA) product. Instead of one fixed screen that every supervisor, agent, analyst, and admin sees in the same shape, the interface is built as "layout-as-data": a set of interchangeable components (page shells, action bars, suggestion chips, forms, charts, artifact canvases, and more) that an orchestration layer assembles differently for each person, based on their role, what they're trying to do, and how they've actually behaved over time.
The system also earns the right to act with increasing independence. Every AI-suggested action starts out requiring a human's approval. As a person approves the same kind of action reliably, that workflow can be promoted — with the person's consent — to "notify only" or fully autonomous execution. Any mistake, any undo, any rejection can instantly demote that autonomy. Nothing about this trust is assumed; all of it is measured.
That measurement is the second half of the project, and it's the larger half of what's in this document: an 88-metric catalogue that watches everything from how someone types a prompt, to whether an approval looks like real supervision or just a rubber stamp, to whether the system is helping different groups of people equally. Nine production algorithms read these metrics and make the interface adapt in real time. A set of foundational component rules makes sure every part of the interface reports what it's doing and never gets stuck in a broken state.
A working prototype (the HTML file attached to this project) demonstrates the idea with four named personas — a supervisor, an agent, an analyst, and an admin — plus a fifth, internal-only "UX Master" view that shows what the analytics catalogue looks like once it's running against real data.
The Problem We're Solving
Enterprise software built for one role tends to fail everyone a little. Concretely, this project exists to solve six specific gaps identified in a review of an earlier, simpler measurement approach:
Sessions were treated as bags of clicks, ignoring the order people actually do things in — so the product never learned the real "happy paths" people discover on their own.
Users were measured only as the persona they were assigned (Supervisor, Agent, Analyst…), ignoring how they actually behave — often 4–6 real behavioural types exist where 3 roles were designed.
Metrics were read one at a time, which misses the interactions between them — a Trust Index can rise at the exact moment real supervision is quietly disappearing, and a single-metric view would never catch it.
Shared, collaborative work (two people looking at the same file) had zero dedicated measurement, even though personalisation changes what "the same file" looks like to each viewer.
The system measured whether AI helps right now, but never whether it's teaching people or making them dependent — a genuine duty-of-care question almost nobody in the market answers.
The whole approach was cost-blind — helpfulness with no accounting for what each helpful action costs in inference spend.
Solving all six required rebuilding the interface itself around adaptivity, and rebuilding measurement into an 88-metric catalog that treats trust, efficiency, quality, and fairness as things to prove continuously — not assume once at launch.
Core Concepts: How the Adaptive UI Actually Works
A screen is represented as a tree of data — an "archetype" (the overall page shape) plus a set of components filling defined slots, each with its own configuration. The renderer's only job is to turn that tree into pixels; it holds no layout opinions of its own. That means the same underlying components can be assembled into very different screens for a supervisor versus an analyst, without anyone hand-coding two separate pages.
The prototype ships with four designed personas (Supervisor, Agent, Analyst, Admin), each with its own default archetype and navigation order. But the catalog also discovers behavioral clusters directly from usage data (metric #69), independent of those assigned roles, and formally tests whether those discovered clusters predict outcomes better than the designed personas do (metric #70). When they do, that's the data nominating a redesign — a human still decides whether to act on it.
Every AI-initiated action is classified by "blast radius" (how much damage a mistake could do) and dispatched through one of three modes: require approval, notify only, or fully autonomous. Promotion between tiers requires a sustained high approval rate with zero critical rejections, and even then the system only offers the promotion — it never silently grants it. Demotion is instant and asymmetric: one undo of an autonomous action, one critical rejection, or a dip in the recent approval rate drops the tier immediately, with a cooldown before it can rise again. An admin kill-switch can freeze autonomy for a whole workflow or the whole tenant at any time.
A small, fixed percentage of users (about 3–5%, assigned once per person and rotated quarterly) always see the plain, static interface, no matter how the adaptive system would have treated them. Comparing outcomes between the adaptive group and this holdout group is the only way the project claims any benefit is real — that comparison is Adaptation Lift (metric #1), and it's checked before any other claim about the system is trusted.
Adaptivity has a cost: an interface that keeps rearranging itself is exhausting, not helpful. Every layout change is priced (a small reorder costs less than swapping in a different component, which costs less than switching the whole page archetype), and each session has a capped "budget" to spend. Once the budget is spent, further changes queue until a natural break point instead of interrupting someone mid-task. If a person manually overrides a layout choice, the system stops making further orchestrated changes for the rest of that session — a courtesy rule.
A handful of metrics are wired as guardrails that can block a rollout or trigger an automatic rollback outright: Adaptation Lift going negative, the undo rate, the escalation rate, and the Cognitive Load Proxy. Two newer, hard-won additions from the v2 catalogue — the Rubber-Stamp Index and a cost spike in AI Economics — are being tracked as future guardrail candidates.
The Prototype: Meet the Users
The attached HTML prototype ("CEA · Adaptive Workspace") is a working, click-through simulation of the ideas above. A person switches between accounts from the top-right avatar menu, and the entire interface — navigation order, page content, even which cards are shown — changes to fit whoever is "logged in." A screen-reader-only heading in the file states the intent plainly: "the interface reshapes itself per person; switch users from the top-right avatar to see the supervisor, agent, analyst, and admin views."
The Decision Layer: Nine Algorithms That Act on These Metrics
Metrics describe what's happening; these nine algorithms are what actually change the product in response. Every one of them serves static, non-adaptive defaults to holdout users, carries a version tag on every decision it makes, and defines its own safe "degrade" behavior for when something upstream fails.
GitHub Repo: https://github.com/gajananrajput21/Sparkathon2026