Contact me
arrow_back

AI Studio · Innovaccer · Apr 2025 – May 2026

Hospital teams build their own AI workflows on it. I led the design.

$11M ARR within 8 months of launch. A true 0 to 1: 14 months from the first sketch.

The bottleneck

We designed every solution by hand. Then the customers multiplied.

By 2025, Innovaccer's data platform could take a new organization's data live in a day. So the organizations multiplied. And every new one arrived with its own workflows, its own edge cases, its own definition of "how we do things here." A product team that designs every solution by hand becomes the bottleneck to its own platform.

The industry's answer to a custom workflow has always been a ticket: a quote of $100,000, a wait of six months, and a feature that's obsolete by the time it ships. With customers multiplying, that math stops working. The custom work had to move out of our hands and into the customer's own.

Custom workflow · prior-auth intake
Quote $100,000 · ETA six months
Queued

Prequel: Data Studio, the decade of data behind this

The users

People who have never written a line of code

The people who needed those custom workflows are front-desk staff, admin workers, care coordinators. Not engineers, not analysts. Their entire technical world is two things: the EHR they were trained on, and a ChatGPT or Claude tab they opened on their own.

That was the design bar for everything that follows: if a surface needs a manual, it is dead on arrival. If it feels less obvious than a chat window, they already have a chat window.

The EHR · trained on it
09:00SMITH, JF/URm 2
09:15OKAFOR, ANEWRm 1
09:30LEE, MF/URm 3
09:45GARCIA, PLAB
10:00PATEL, RF/URm 2
10:15NGUYEN, TNEWRm 1
6 OF 42 · TODAY · PAGE 1/7
FDCCAD
A chat tab · opened on their own
You: what does prior auth status "pending review" mean??
Copilot: It usually means the insurer has your request but hasn't assigned a reviewer yet…
Their entire technical world · we had to live in the gap

The bet

Ship the generic platform, let the hospital build the last piece

March 2025. Innovaccer set a new direction, called Gravity inside the company: turn its healthcare intelligence into something users build with, not just use. Two products, one bet. The Enterprise Copilot is the promise: an AI that does the work you ask of it. AI Studio is the factory: the place where the hospital's own people assemble that intelligence into their own workflow.

The room: my design manager, a design peer, product and engineering leaders, three PMs, seven developers, and me. I led the design. Fourteen months from the first sketch.

Final assembly, by the customer

Payer policiesKnowledge base
Prior-auth checkerAgent
Intake triageAgent
PHI guardGuardrail
Eligibility APITool
Referrals, end to endWorkflow
The generic platform
Talk to my Portfolio
Try one:
⌘ K

Live: answers come from a curated knowledge base, behind guardrails. Nothing you type is stored. This is inspired from the Gravity Enterprise Copilot

Scene 1

The argument I lost

Before a single screen of AI Studio existed, we had to answer one question: build the agentic backend ourselves, or ship the first version on existing AWS and Azure APIs while our own backend caught up behind it. I argued for building our own, and I was not quiet about it.

Read the scene: build it or borrow it

Reconstructed from memory. The words are close, the decision is exact.

Me: If we ship on their APIs, our v1 is every other AI studio in the market. Same capabilities, same ceilings. Where's the healthcare intelligence in that?

My PM: It's not in v1. It's in what v1 teaches us. We're new to this category. Every feature we build from scratch right now is a guess.

Our engineering lead: An expensive guess. Their APIs put us in front of real users this quarter. We roll out our own backend behind it, piece by piece, wherever we hit a wall.

Me: So instead of building features on a whim, we wait for the feedback to tell us what to build.

My PM: Exactly. And the experience stays ours. The design quality doesn't drop because the engine is borrowed.

That last line settled it for me. I could live with a borrowed engine. I could not have lived with a borrowed experience. They were right, and the proof came fast. Within six months we were replacing pieces of the borrowed backend one by one, exactly where real usage hit real walls, and we moved to LangSuite for the broader capabilities. Not one of those rebuilds was a guess.

I lost that argument. The users won it.

Launch · all borrowedSix months · rebuilt where usage hit walls

The centerpiece

One canvas for the whole build

Every capability in AI Studio works as a standalone thing. An agent, a knowledge base, a guardrail, a workflow. But agents need knowledge bases, guardrails, tools, sub-agents. Workflows need all of them. So the canvas had to hold every layer at once and let you move between them without losing the plot. That surface exists because I pushed for it, and it's the thing I'd defend the hardest.

This is the bet made real: the generic platform underneath, the hospital's own use-case assembled on top, by the same people who used to file the ticket.

Build · Code · Monitor · every layer on one surface

Reconstructed · drag nodes · drag a port to connect · click to edit · del to remove

Referrals workflowDraft
Use to add nodes in place
all_inclusivePATIENT LOOPcollapse_content
play_arrow
account_treePatient intakeCollects the referral and cleans it up
robot_2Prior-auth checkerChecks coverage against payer policies
alt_routeCheck decisionApproved, or needs review
robot_2Coordinator handoffPrepares the case for human review
buildNotify clinicPosts the decision back
logout

Scene 2

We killed the Python notebooks

The ML story was half-designed: Python notebooks, proper data science tooling, the works. Customer excitement was with AI, and in a category this new you follow the pull. So we killed it. Right call? I think so. Ask me again in a year.

ml-studio.ipynb · unshipped
[1]
[2]
Run all
SDOH notebookKernel off

Do comorbidities track social risk?

Care coordinators keep seeing it: patients with tougher life circumstances carry more diagnoses. This notebook checks that hunch on a synthetic cohort (no real patients; seeded, so the numbers reproduce).

  1. Build 120 synthetic patients with an SDOH risk score (0 to 100) and a comorbidity count.
  2. Fit a least-squares regression line.
  3. Plot the cohort, the fit, and the correlation.

If the slope holds, rising social risk means rising clinical complexity: exactly the patients referrals should reach first.

The one we killed. Real Python, running in your browser

Scene 3

Tested like a patient's life depends on it

A normal AI product may or may not have evals. In healthcare, evals are the difference between a real patient's life and death. That is not a line I use to sound serious. It is the actual design requirement.

The decision I am proudest of: evals and datasets are curated per use-case, so the front-desk coordinator assembling her workflow never sees a rubric, a dataset, or the word "eval." The machinery is healthcare-grade precisely so she never has to feel it.

Payer policiesLive
Knowledge base indexing completed.
Before agents can use it, it needs to be evaluated.
Basic details
DescriptionEvery payer's prior-auth and coverage policies, split by clause, so the checker can cite the exact line.
Datasets
NameChunkingSourceStatus
Commercial payersClauseSFTP dropIndexed
Medicare LCDsClausePortal scrapeIndexed
Denial letters Q3HierarchicalObject storageIndexed
Embedding details
ModelTEXT-EMBED-HC-2
TypeFloating-point vector embeddings
Vector dimensions1024
Auto-indexingNightly
Storage details
Vector storeIn-house · Gravity
Namespacepayer_policies
Knowledge graph

Scene 4

A chat that answers with the product, not a paragraph

Everyone's first instinct for a copilot is a chat sidebar. We said no. Side panels split your focus across surfaces, they don't translate to mobile, and they make the AI feel like an assistant hovering at your elbow. We built a single panel where chat and widgets live together: the smallest piece of product interface that can appear inside a conversation.

Remember the users: half their technical world is already a chat window. The Copilot meets them at the one AI interface they already trust. A collaborator sitting next to you, not a helper waiting outside.

You: What needs my attention before ’s 2:00 visit?
Copilot: One thing that should not wait. Her MRI prior-auth cleared overnight, and she should hear it before her visit so imaging can be scheduled while she is here. I can call her now and handle it. You will only be pulled in if she asks something I cannot answer.
LJ
Lawson, Joy
+1 (856) 986-1765
Prior-auth follow-up

Play the coordinator: start the call and follow it to a printed referral

The receipt

Nothing gets sold without the Copilot

I wasn't in the sales calls. That one line is what came back from the people who were. AI Studio launched in May 2025; by January 2026, 8 months in, it had crossed $11M ARR. Xccelerate, Innovaccer's annual conference, carried the design vision on stage in '25 and '26.

Closing

Users build their own solutions now

Back then I believed I was designing stuff for our users. Now I believe I'm enabling them to design and build solutions on their own.

None of these screens are product screenshots. Everything here is rebuilt in my own design library, iterated past what shipped. NDAs are a constraint. Constraints are a brief.

Companion read

Stop Shipping "Finished" AI is the same thesis, derived from a 1990s printer. What HP's DeskJet supply chain knew about healthcare AI before healthcare AI existed.

Prequel: Data Studio, the decade of data behind this