Midstream Health | rohremala

<- Midstream Health

Designing an AI-Driven Healthcare Procurement Platform

Team

Rohitha Remala,

Angela Chavez-Luna

Kristine Mudd

Roles

Overview

Healthcare procurement teams operate in an environment defined by:

Fragmented contract and rebate data
Manual reconciliation across spreadsheets and vendors
High financial and compliance risk
Deep distrust of “black box” automation

The challenge was designing an AI-enabled system that:

Increased transparency rather than obscuring logic
Augmented human decision-making instead of replacing it
Could scale across multiple procurement verticals
Remained trustworthy even when AI outputs were probabilistic or incomplete

This required designing not just UI—but confidence, explainability, and operational trust.

The Core Thesis

Healthcare procurement teams operate in an environment defined by:

Fragmented contract and rebate data
Manual reconciliation across spreadsheets and vendors
High financial and compliance risk
Deep distrust of “black box” automation

The challenge was designing an AI-enabled system that:

Increased transparency rather than obscuring logic
Augmented human decision-making instead of replacing it
Could scale across multiple procurement verticals
Remained trustworthy even when AI outputs were probabilistic or incomplete

This required designing not just UI—but confidence, explainability, and operational trust.

Approach

My Mandate

As the first and only designer, I owned:

Defining the product UX strategy from zero
Translating AI and data capabilities into usable, explainable workflows
Establishing the design system and UI architecture
Acting as the bridge between product, engineering, data, and sales
Designing for future scale across healthcare verticals

This was a systems ownership role, not a feature delivery role.

Objective

Rather than designing features upfront, I focused on reducing uncertainty through structure.

1. Mapping Reality Before Designing Screens
I worked with stakeholders to model how sourcing, rebates, and contracts actually flowed across teams—surfacing decision points, failure modes, and trust gaps before committing to UI.

2. Treating AI as Visible Assistance, Not Hidden Automation
AI insights were intentionally designed to be:

Inspectable
Traceable
Bounded

Users needed to understand why the system surfaced an insight—not just what it suggested.

3. Using Prototypes as Alignment Infrastructure
High-fidelity, end-to-end prototypes became the primary artifact for aligning leadership, engineering, and sales around what the product could and could not responsibly do.

Engineering-Aware Design

RAG as a First-Class Constraint

A defining complexity of this product was that every major UX decision depended on backend AI systems, specifically Retrieval-Augmented Generation (RAG) pipelines owned by the data engineering team.

This was not a scenario where design could operate independently of implementation. The usability, credibility, and compliance posture of the product depended on tight alignment between:

Frontend UI states
RAG retrieval reliability
Model confidence and latency constraints

Designing Against Real AI Capabilities (Not Assumptions)

Each workflow I designed explicitly accounted for:

What data could be retrieved reliably vs. probabilistically
Latency variability in RAG responses
Partial, missing, or conflicting source data
Traceability requirements for healthcare procurement decisions

Rather than designing “ideal” AI interactions, I treated RAG outputs as a system dependency, shaping the UI to surface:

Confidence levels and source provenance
Degraded and fallback states
Clear boundaries between AI-assisted insight and human judgment

This ensured the product remained operationally usable even when AI outputs were imperfect.

Frontend–Backend Contracts as a Design Responsibility

I treated frontend–backend alignment as a design problem, not an engineering afterthought.
For every major feature, I partnered with data and engineering to define RAG output schemas, validate which UI states were consistently supportable, and align on error, latency, and fallback behaviors. No design was considered final until feasibility was signed off and empty, error, and degraded states were fully designed.

This approach prevented downstream rework and avoided UI promises the system could not reliably fulfill.

Down the road, this same contract-first thinking directly led to the creation of an internal data validation tool used by staff to inspect RAG outputs before they surfaced in the product. The tool allowed internal teams to:

Validate source coverage and retrieval quality
Review confidence signals and failure cases
Catch data inconsistencies before they reached end users

By extending design ownership upstream into internal tooling, we reduced production risk, improved AI output quality, and created a shared operational language between design, data, and engineering—reinforcing trust both inside the organization and in the customer-facing product.

Designing for Explain-ability & Trust in AI Outputs

Because procurement decisions carry financial and compliance risk, AI explain-ability was non-negotiable. I intentionally designed interfaces that:

Made AI insights inspectable rather than opaque
Clearly separated retrieved facts from inferred recommendations
Allowed users to understand why an insight surfaced

This required tight iteration with data engineers to ensure RAG outputs were accurately attributed to their source data, structured in a way the UI could reliably consume, and paired with meaningful confidence signals that users could understand and trust when making high-stakes procurement decisions.

Product Build

Phase 1 — Core Rebates Platform

Defined foundational IA, end-to-end rebate workflows, and data visualization patterns that became the interaction grammar for the platform.

Phase 2 — Service Layer Expansion (Email Intelligence)

Designed an email-based insight service to extend AI value beyond dashboards and into real procurement workflows.

Phase 3 — Build & Engineering Partnership

Worked through design QA, feasibility constraints, and UX integrity during implementation.

Phase 4 — Research-Driven Expansion

Led research to identify adjacent workflows, informing expansion into Managed Care and Pharmacy verticals.

Phase 5 — Scaling Through Systems

Built the foundational design system:

Tokenized styles mapped to root CSS variables
Scalable components and interaction patterns
A design system CMS for documentation and versioning

Designing for Sales, Not Just Users

In enterprise healthcare, buying decisions precede usage.
I designed fully interactive, product-grade sales demos in Figma that:

Simulated real data and workflows
Demonstrated AI-assisted decision points
Aligned sales narratives with actual system behavior

These demos shortened sales cycles, reduced engineering dependency, and aligned internal teams around a shared product truth.

Brand as a System Signal

Scope

As Midstream evolved from an early-stage prototype into an enterprise-facing healthcare platform, its existing branding no longer reflected the maturity, credibility, and trust required for regulated, investment-heavy procurement workflows.

What I worked on:

Brand Identity refresh (took a part)
Landing page redesign
UI & UX (Core screens)
Design system development
Graphic Icons

Design Direction

Refreshed the brand to reflect a mature, enterprise healthcare platform and support future expansion into Managed Care and Pharmacy.

Kept the interface data-forward and restrained, prioritizing clarity and trust over decorative visuals.

Built reusable, systemized UI components to support scalable workflows and new clients.

Designed landing and entry experiences to clearly communicate Midstream’s value, scope, and AI-driven capabilities.

Concepts

Outcomes, Impact & Ownership

Shipped Midstream’s first production AI healthcare product
Established a scalable UX and design system adopted across teams
Enabled multi-vertical expansion (Rebates, Managed Care, Pharmacy)
Accelerated sales with product-grade interactive demos
Created a design foundation capable of supporting future AI evolution

The impact of this work came from designing under extreme product and AI ambiguity, owning systems rather than surfaces, translating probabilistic AI into human trust, and treating engineering constraints as core design inputs. Success was not measured by screens shipped, but by clarity created, systems scaled, and risk reduced across product, data, engineering, and sales.