Lokavant case study · Clinical-trial intelligence

Making AI clinical forecasts usable, explainable, and operational.

I reworked dense CRA and CRO workflows during a platform consolidation, then connected product strategy, UX, design systems, front-end implementation, demos, and marketing into one reusable system.

Role Principal Product Designer

Scope Product strategy, research, UX/UI, design systems, front-end, demos, and GTM

Collaboration Product, clinical, data science, engineering, sales, and marketing

Selected work 2024–2025

Demo production 3 weeks → 3 days

A reusable product and storytelling system replaced one-off production.

LinkedIn CTR 12% → 37%

Sharper product positioning improved campaign response.

Platform data 500K+

Studies represented across integrated data sources.

Forecast generation <5 min

Product capability communicated in shipped materials.

Lokavant Spectrum workspace with an enrollment forecast, country activation view, scenario controls, and a table of forecasts — Spectrum brought forecasting, scenario comparison, optimization, protocol intelligence, and the forecast library into one study-planning workspace.

Problem / Powerful models, fragmented product

The forecasts were sophisticated. The decisions around them were harder than they needed to be.

Clinical teams moved between dense tables, charts, wizards, and disconnected tools to plan enrollment, compare scenarios, and adjust active studies. Important assumptions and system states were difficult to follow.

A forecast is not a single prediction. Teams must define the study, understand protocol constraints, choose countries and sites, model enrollment, estimate cost, compare historical evidence, and revisit the plan as real data arrives. The interface exposed pieces of that machinery without yet giving users one coherent decision path.

Development-state Lokavant create-forecast screen dominated by a country chart and raw study table without a clear staged workflow — A development-state forecast screen showed the issue clearly: dense source data was visible, but the user’s next decision and the path through the work were not.

Operational problem

CRAs and CRO teams needed to move between planning and in-flight study management without losing the assumptions, evidence, or reasoning behind a forecast.

Trust problem

AI outputs could look authoritative while hiding uncertainty. Teams needed to see confidence ranges, comparisons, constraints, and consequences before acting.

Why it mattered / Decisions under pressure

A forecast only matters if a team can understand and defend the decision behind it.

Country selection, site mix, cost, target enrollment, and last-patient-in timing affect one another. A small change can alter feasibility, budget, and delivery risk across an entire study.

Plan variables

Countries Sites Protocol Enrollment assumptions

Model response

Feasibility Confidence range Cost Target LPI

Team decision

Proceed Compare Optimize Reforecast

Change one input and the downstream decision can move. The interface therefore had to expose dependencies and consequences—not simply report a model output.

Approach / Reframe the product

Organize the platform around decisions, not around models.

I reframed Spectrum as a trial decision workspace. The product needed to help teams predict what might happen, optimize a plan against real constraints, and control the study as conditions changed.

Predict

Make enrollment trajectory, target LPI, country activation, confidence, and odds of success understandable at a glance.

Optimize

Let teams compare country and site plans, costs, time constraints, and feasibility without obscuring the tradeoffs.

Control

Bring actual study data back into the forecast so teams could replan without starting over or losing context.

Spectrum diagram connecting public, proprietary, and licensed data across more than 500,000 studies to forecasting and country-and-site optimization workflows — The product connected a large clinical-trial data foundation to forecasting and optimization workflows. My design work focused on making that analytical capability usable as a sequence of operational decisions.

Execution / Building a forecast

Translate model requirements into a staged workflow.

The workflow moved from study definition through country and site assumptions, cost, protocol intelligence, and comparable studies. Users could start from an existing study or run a pre-study scenario, while the product preserved progress and made generation status visible.

Forecast creation workflow beginning with forecast information and the choice between an existing study and a pre-study scenario — 1. Define the forecast and choose the planning context.

Forecast workflow for country and site assumptions, cost inputs, and protocol intelligence — 2. Add country, site, cost, and protocol assumptions without flattening the complexity.

Forecast generation dialog and forecast library with running, completed, and failed states — 3. Generate the forecast, keep running, completed, and failed states visible, and return users to a library they can compare.

Execution / Making uncertainty legible

Separate what happened from what the model predicts.

Trust required more than a forecast line. The interaction distinguished actual enrollment, predicted performance, confidence bounds, and the original plan, then connected the trajectory to country-level regulatory and site-activation activity.

Reforecast visualization combining actual enrollment with a forward prediction range and country-level site activity — Reforecasting combined observed enrollment with a forward confidence range and country-level site activity so teams could judge whether the plan still supported the target.

Execution / Make tradeoffs explicit

Show what optimization changes—and what it costs.

Optimization could not behave like a black-box recommendation. Results exposed the cost difference, LPI difference, enrollment trajectory, and operational assumptions behind the proposed plan.

Optimized Lokavant forecast showing a 31-day LPI difference, a cost difference, enrollment prediction, confidence range, and operational metrics — An optimized forecast connected the recommendation to its cost, timeline difference, enrollment trajectory, and underlying assumptions.

What shipped

One system across product, demos, and go-to-market.

The work extended from production workflows and reusable components into the demos, visual language, website, video, and social content used to explain the product.

Forecasting

Study setup, enrollment projections, confidence ranges, comparator evidence, scenarios, cloning, and reforecasting.

Optimization

Country and site planning, candidate-site ranking, weights, cost constraints, time constraints, and explicit tradeoffs.

Reusable system

Shared design language and components connecting Figma, production UI, front-end implementation, and sales demos.

Product story

Positioning, website, videos, social campaigns, and editorial content grounded in the same product decisions.

Production Lokavant forecast detail with enrollment prediction, country activation, odds of success, forecast parameters, and costs — The production forecast workspace connected the enrollment trajectory with odds of success, country activity, assumptions, costs, and detailed study evidence.

Lokavant candidate-site recommendations with configurable weighting, scores, study experience, proximity, and enrollment rates — Candidate-site recommendations made the ranking factors visible and adjustable rather than presenting an unexplained score.

Lokavant Spectrum website presenting the promise of forecasting a clinical trial in minutes rather than weeks, with product visuals and video — The same decision-centered story shaped the website, demos, video, and editorial system: forecast at the site level, adjust in real time, quantify uncertainty, and continuously reforecast.

Impact

Faster production, clearer positioning, and a product story that held together.

3 weeks → 3 days Reusable demo production

12% → 37% LinkedIn click-through rate

<5 min Forecast-generation capability communicated in Spectrum materials

500K+ Studies represented across integrated platform data

Demo-production and campaign metrics describe the reusable design and go-to-market system. The forecast-generation and study-count figures are Lokavant product and platform claims presented in shipped materials. Product, clinical, data science, engineering, sales, and marketing contributed to the product and business outcomes.

What it proved

AI products earn trust through visible reasoning.

The strongest design work was not making the model appear intelligent. It was making the evidence, uncertainty, constraints, and consequences clear enough for people to act.

Trust is interaction design

Confidence ranges, comparisons, assumptions, system states, and explicit tradeoffs make an AI recommendation inspectable rather than magical.

Design decisions, not dashboards

Complex enterprise products become usable when information architecture follows the work people must complete—not the shape of the underlying data.

One story should cross the company

Product UX, engineering, demos, positioning, and marketing become faster and more credible when they share the same language and system.