Infinitus AI · Healthcare · 2023–Present

Studio

The design platform for Infinitus, where healthcare teams build, test, optimize, and deploy the autonomous AI agents that hold real conversations with patients, providers, and payors. What used to take 35–50 pages of prompts and a team of engineers now happens in a single canvas, accelerated by an embedded AI assistant.

My Role

Staff Product Designer & Design Manager

Team

Product, Engineering, Data Science, Clinical, Compliance

Methods

Platform strategy, generative research, system design, AI-native prototyping

The Problem

Healthcare AI was forcing a tradeoff no one could accept

Every enterprise buyer asking about AI governance is really asking about one tension. Fully deterministic agents (scripted call flows) are safe and auditable, but break the moment a conversation goes off-script, which is always. Fully non-deterministic LLMs feel natural, but hallucinate on the exact clinical details that can’t be wrong: drug safety information, benefit disclosures, adverse event responses. Neither is acceptable in regulated pharma.

The teams building these agents felt the same friction. A single program meant wrangling 35–50 pages of prompts across 10+ sub-agents, each with its own precision-tuned configuration. Every new use case got rebuilt from scratch: new connectors, new guidelines, new quality criteria. Insights from production calls were captured by analytics, but nobody could act on them without a data analyst or a months-long IT ticket.

Studio is the platform we built to dissolve both tensions in one place.

Process

From a no-code call-flow builder to a healthcare-grade agent platform

V1 MVP
Learn from operators
Re-architect
Embed the Assistant
Ship & iterate
V1 · 2023 Call Flows: the no-code MVP for HRA-style programs

Studio started as a different bet entirely. Before pharma support programs were on the roadmap, we were building a no-code call-flow tool for non-technical healthcare operators running simpler use cases: patient-facing HRA (Health Risk Assessment) calls for Medicare plans like Zing Health. The premise: anyone with domain knowledge should be able to drag-and-drop a phone call together.

Starting a new flow: template or blank canvas

Block canvas: Guidance, Ask & collect, Condition, Hangup

Live test panel: call your draft, see what fires

Auto-generated Q&A pairs from uploaded clinical content: the first place AI carried real weight in the design surface

What V1 got right. The block-based canvas was the right metaphor for a conversation. The live test panel (call your draft and watch it fire turn by turn) became a non-negotiable in every version after. And the Q&A generator gave us our first proof that AI could carry real weight in the design surface, not just sit beside it.

Where it hit its ceiling. When enterprise pharma showed up needing patient support programs that ran across multiple agents, channels, and weeks of patient journey, the cracks showed fast:

  • Drag-and-drop nodes didn’t scale to multi-agent journeys with branching, retries, and handoffs across days.
  • The fixed step types (Guidance, Ask & collect, Condition, Hangup) forced clinical logic into UI primitives that didn’t match how operations teams actually think about programs.
  • There was no way to govern what the agent could say in critical moments, only sequence and conditions. That wasn’t acceptable for drug information, benefit disclosures, or adverse event responses.
  • No optimization at scale. Every change required hand-editing prompts and re-testing manually.

We rebuilt from those lessons. I led the re-architecture from a single canvas tool into a full ecosystem (Agents, Journeys, Connectors, Lens criteria, Copilots) orchestrated through an embedded AI Assistant that turned the V1 drag-and-drop into a conversation. The mental model became: describe what the agent should do, the system handles the prompt engineering.

AI Assistant one conversational orchestrator across every primitive below

AI Agents

Voice, email, SMS, and chat agents composed from Flow, Topics, and Guidelines.

Journeys

Multi-step workflows orchestrated across agents, channels, and weeks of patient time.

Connectors

Salesforce, EHRs, and any OpenAPI surface, built once and reused by every program.

Lens

Quality criteria that score every production call and surface risk automatically.

Copilots

Hand moments that need human judgment back to people, with full context attached.

Design → See → Improve — production insights return to Studio as recommended next actions
The re-architecture in one picture: five reusable primitives, one conversational orchestrator, and a feedback loop that closes without an analyst in the middle.

Solution

One canvas. Idea to launch, fast.

Today Studio: the design platform for the full Infinitus ecosystem

Studio takes a team from first draft to production-ready in four steps, all in the same place. The same connectors, guidelines, and quality framework get reused across every agent; nothing is rebuilt for the next program.

Agents library: the top-level inventory across every customer program

01 · Build

Design every part of the ecosystem in one place: Agent Connectors for Salesforce, EHRs, and any OpenAPI surface; AI Agents for voice, email, SMS, and chat; Journeys that orchestrate multi-step workflows; Lens criteria that score quality and surface risk; and Copilots that hand work back to humans with the right context. Conversation paths get composed from three primitives we defined: Flow, Topics, and Guidelines.

Every new agent starts with a choice that didn’t exist in V1: Start with AI Assistant. Describe what you want the agent to do in plain language and Studio generates the topics, guidelines, and connector suggestions for you to refine.

Create new agent: three on-ramps, one shared model

AI Assistant building Topics, Guidelines, and Evaluators in-canvas

The Build canvas: topics on the main surface, Assistant on the left, configuration on the right

Agent Response Control · the patented core

ARC is the answer to the tradeoff this case study opened with: scripted agents that break the moment a conversation goes off-script, or fluent LLMs that hallucinate on the details that can’t be wrong. Instead of choosing one posture for the whole platform, ARC decides per moment. Before the agent speaks, it identifies which response type the guideline requires, so governance stops being a platform-wide compromise and becomes a design decision made at the level of a single conversational beat.

“Is it safe to take this with ibuprofen?”

One patient question · three governance modes

Strict

Approved language, verbatim

The agent reads the approved drug-interaction disclosure word for word. Drug safety information, benefit disclosures, and adverse event responses never vary, ever.

Hybrid

Natural phrasing, bounded knowledge

The agent answers conversationally but can only draw from the program’s vetted clinical content. It stays inside the knowledge boundary or escalates.

Fluid

Free conversation, low stakes

Scheduling, wayfinding, and small talk generate naturally, keeping the conversation human where nothing clinical is on the line.

Same conversation, three governance modes, one consistent result: safe, accurate, natural.

Topic detail: what the agent knows how to handle

Guidelines: per-moment When/Then with ARC response type and criticality

02 · Test · 03 · Optimize

Test. Run live test calls right inside the canvas. A turn-by-turn trace shows which guideline fired, which connector ran, and exactly what data moved, the same view the operator gets in production.

Optimize. Instead of editing the 35–50 pages of prompts and sub-agent configurations by hand, you describe intent in plain language. Optimizer generates prompts and settings, simulates thousands of conversations at scale, and scores them against quality criteria. If the score falls short, the loop refines and retests until the agent is ready, about 90% faster than manual tuning.

Agent overview: AI Assistant guides into a live test session the moment configuration is ready

04 · Deploy

Deploy. Ship in one step with dashboards, versioning with rollback, topic clusters, and agent health monitoring live from day one. Issues get flagged automatically and feed back into Studio as recommended next actions, closing the Design → See → Improve loop without an analyst in the middle.

Tasks: every production run, monitored from day one

Lens Insights: production patterns surfaced as recommended actions back in Studio

Outcomes

Measurable impact

~90%

Faster than manual prompt tuning

100%

Of calls monitored from day one (vs. 2–5% manual audits)

90% / 0%

Patient-rated positive interactions with zero compliance failures

Days

To connect a new EHR, CRM, or payor system, not months

Qualitative

  • Collapsed the build cycle: programs that used to require eng-led prompt engineering ship from a single canvas, with reusable connectors, guidelines, and quality criteria across every agent.
  • Made governance designable: ARC turned the safety-vs.-naturalness tradeoff into a per-moment design decision instead of a platform-wide compromise. The result is conversations enterprise compliance teams trust and patients actually like.
  • Removed the analyst bottleneck: Infinitus Assistant lets operators ask “what changed this week?” or “why is the greeting under-performing?” and apply the fix from the same conversation. No IT ticket required.
  • Compounded across customers: every agent built in Studio reuses the same connectors and quality framework, so each new program ships faster than the last.

Next

Where Studio goes from here

Now In active planning
  • Assistant as operator, not just helper. Lens Insights stop being read-only recommendations. They become proposed guideline diffs and Optimizer runs that the Assistant drafts and a human reviews-and-approves before deploy. Hand-tuning becomes the exception, not the default loop.
Then Building toward
  • Self-serve Studio for customers. Pharma ops teams author and tune their own agents inside a sandboxed surface, with shared ARC governance and the same Lens criteria the Infinitus team uses. Customers move at the speed of their own program decisions instead of waiting on us.
Longer horizon Bets on the platform’s compounding value
  • A marketplace of reusable building blocks. Agents, Journeys, Connectors, Lens criteria, and full program templates published and versioned across customers, so each new program assembles mostly from proven parts.
  • Multi-agent journeys across longer patient timelines. Shared state and memory across days and weeks, agent-to-agent handoffs, and expansion into video, embedded EHR/CRM SDKs, and other surfaces patients already live in.