← Case studies

Virtova case study · Healthcare services — health-insurance communications · On-premises AI member-engagement platform deployment (Limni)

May 6, 2026 · By Sultan Meghji

Healthcare member engagement, $7–10M projected, no record off-premises

Projected digital revenue

$7–10M

Member records analyzed

1.7M+

Data access → first recs

10 days

Actionable segments

5

Disclosure. Limni is built by Frontier Foundry Corporation. Sultan Meghji is the founder of Virtova LLC and the Co-Founder and CEO of Frontier Foundry; the two firms share ownership. Virtova was the advisory layer on this engagement; the Frontier Foundry platform was selected on the merits against the firm’s PHI-adjacency, data-rights, and insurer contractual constraints after a vendor-elimination diligence pass. The relationship is also documented at /editorial/.

The engagement at a glance

A leading U.S. healthcare-services organization manages member communications, benefit activation, and portal onboarding for a top-tier national health insurer covering multiple millions of insureds. The historical outreach model was print-led: portal-registration was approximately 2.5% of the eligible member base, and additional print volume was producing diminishing conversion returns at rising cost.

Virtova led the program. Frontier Foundry built and deployed the on-premises member-engagement intelligence platform on Limni. Ten days from first data access to first channel recommendations. Approximately six weeks to full operational deployment. Five distinct member segments, each mapped to a highest-yield outreach channel. $7–10M in projected digital-engagement revenue. No member record left the client’s infrastructure.

The three structural constraints

Untargeted outreach at population scale. Campaigns targeted broad demographic slices rather than members statistically likely to respond. The result was low conversion, unnecessary print volume, and preventable call-center load — three costs the firm was carrying for every campaign cycle regardless of the response rate.

Sparse source data with no behavioral signal. Starting datasets contained only address, ZIP code, and date of birth per member. None of the standard inputs that drive consumer-engagement modeling. The first half of the engagement was about turning that into a usable feature set without crossing any data-rights boundary.

Cloud AI was off the table. Commercial AI providers were disqualified on data-security grounds. Transmitting PHI-adjacent member records to an external inference provider would have violated the firm’s regulatory posture and its contractual obligations to the underlying insurer. The architecture had to be on-premises before any modeling decision was made.

The Virtova advisory layer

Virtova ran the program. The advisory work covered four threads that have to land before a system like this ships into a healthcare-communications firm:

  1. Data-rights and compliance posture. Member data classification, PHI-adjacency analysis, contractual obligations to the insurer mapped against the data flow. The engagement could only proceed once a written data-rights memo was acknowledged by the firm’s CCO and reviewed by the insurer’s audit team.
  2. Use-case scoping and prioritization. Engagement was sized to a single high-leverage outcome — member segmentation driving channel selection — rather than a broad AI-everywhere mandate. The exit criteria were defined before the first model trained.
  3. Vendor-elimination diligence. Cloud vendors were disqualified on data-residency. Self-hosted vendors were evaluated against actual model capability over the firm’s data shape. Limni was selected because it was the only architecture that combined on-premises LLM inference with the explainability the firm needed to defend its outreach decisions to the insurer.
  4. Operating handoff. The platform went live with a documented run book, a human-review surface for every recommendation, and a quarterly performance review against the metrics in this engagement.

The Limni deployment

Limni is Frontier Foundry’s proprietary LLM infrastructure, deployed as a self-hosted model trained on the client’s own data and inferring exclusively inside the client’s perimeter. Architecture and capabilities deployed at the firm:

What changed in production

The $7–10M projected digital-engagement revenue is the operating-leverage story. Converting unregistered members through targeted digital channels at rates materially above the 2.5% baseline produces revenue per converted member that compounds against the cost of the prior print-only model.

Five actionable segments replaced an undifferentiated demographic-slice campaign model. Each segment has a defined channel strategy, a documented predictive signal, and an explainable recommendation the firm can defend to the insurer’s audit team.

The synthetic data generator changed the engagement’s tempo. Once analysts could iterate on workflow design against PHI-free data that preserved statistical structure, the friction of “schedule a data-rights review for the next test” disappeared. That single artifact saved weeks across the lifecycle.

The platform continues to surface macro-level trends in digital engagement, letting the firm reallocate spend from low-conversion paper outreach toward the higher-yield digital channels per segment. The system stays inside the perimeter.

Why this combination of work is rare

A regulated healthcare-services firm with sparse source data, PHI-adjacency constraints, contractual obligations to a top-tier insurer, and a print-dominant outreach legacy is the exact problem most AI vendors cannot solve. The cloud vendors cannot meet the data-residency posture. The on-prem vendors usually cannot ship a working LLM. The boutiques that can do both rarely have a regulated-industry program management discipline strong enough to keep the engagement on a defensible compliance footing through delivery.

Virtova plus Frontier Foundry sit at that intersection — the advisory layer that makes the platform legible to the firm and to the insurer, and the platform layer that meets the data-rights constraints.

Engagement model

Bespoke, fixed-scope phase one. The engagement opens with a confidential scoping call against the client’s data shape, compliance posture, and target outcome. Phase one closes at first operational deployment with the measurement baseline in place. Phase two is operations.

Book a discovery call →

"PHI-adjacent data does not get sent to an external inference provider. Not for cost. Not for convenience. The architecture has to be on-premises before anyone draws a model on a whiteboard. Everything downstream is conditional on that."
— Sultan Meghji

Engagement model

Frequently asked

About this engagement

What outcomes did the engagement deliver?
Five actionable member segments with documented predictive signals and channel-specific outreach guidance; a channel-prioritization framework letting the firm shift spend from low-conversion print to higher-yield digital channels; $7–10M in projected digital-engagement revenue; 1.7M+ member records analyzed in-perimeter. Ten days from first data access to first channel recommendations; approximately six weeks to full operational deployment.
Why was cloud AI ruled out for this firm?
Commercial AI providers were disqualified on data-security grounds. Transmitting PHI-adjacent member records to an external inference provider would have violated the firm's regulatory posture and its contractual obligations to the underlying insurer. The architecture had to be on-premises before any modeling decision was made.
What did the Virtova advisory layer cover?
Four threads: data-rights and compliance posture (PHI-adjacency analysis, written memo acknowledged by the firm's CCO and reviewed by the insurer's audit team); use-case scoping (sized to a single high-leverage outcome with defined exit criteria); vendor-elimination diligence (cloud vendors disqualified on data-residency, self-hosted vendors evaluated against the firm's data shape); operating handoff (documented run book, human-review surface for every recommendation, quarterly performance review).
What platform components were deployed?
Kundi secure data layer (all ingestion, cleansing, enrichment, training, and inference flowed through on-premises storage); significance-tested feature selection via logistic regression; K-means member segmentation producing five segments; Limni-powered recommendation LLM trained on the client's data and inferring exclusively inside the perimeter; synthetic data generator producing statistically identical PHI-free datasets for analyst workflow testing; channel-prioritization framework mapping each segment to highest-yield outreach channels.
How is this engagement different from typical AI consulting?
A regulated healthcare-services firm with sparse source data, PHI-adjacency constraints, contractual obligations to a top-tier insurer, and a print-dominant outreach legacy is the exact problem most AI vendors cannot solve. Cloud vendors cannot meet the data-residency posture. On-prem vendors usually cannot ship a working LLM. The boutiques that can do both rarely have the regulated-industry program management discipline. Virtova plus Frontier Foundry sit at that intersection.
What was the synthetic data generator and why did it matter?
A statistically identical, PHI-free dataset that preserved the structure of the real member data. Once analysts could iterate on workflow design against synthetic data, the friction of "schedule a data-rights review for the next test" disappeared. That single artifact saved weeks across the lifecycle and is one of the highest-leverage tools the engagement produced.

Work with Virtova

Most engagements start with a 30-minute call.

Confidential by default. NDAs available on request.

Book a discovery call →