Context
SameSky Health is a health equity company that helps health plans engage underserved populations through culturally-sensitive, personalized outreach. The core insight behind the company: traditional member engagement (generic mailers, automated phone trees, one-size-fits-all portals) fails the people who need it most. Members from diverse cultural backgrounds, non-English-speaking households, and historically marginalized communities disengage from healthcare systems not because they don't care about their health, but because the system doesn't speak to them in ways that feel relevant, respectful, or trustworthy.
When I joined as VP of Product, the company was early-stage with a vision but no scalable platform. My job was to build it.
The Real Question
The surface-level challenge was straightforward: build a platform that delivers personalized health engagement content to members. But the real design problem was far more complex.
This was a multi-sided marketplace with three distinct stakeholders, each with different needs, constraints, and definitions of success:
The payer (health plan)
Needs to improve quality measures (HEDIS, Stars ratings), reduce cost of care, and demonstrate health equity impact. Pays for the service. Measures success in gap closure rates and member activation.
The member
Needs health information and motivation that feels personally relevant, culturally appropriate, and actionable. Didn't ask for the outreach. Measures success in whether their actual health needs are being addressed.
The community health worker (CHW)
Needs to build trust with members, deliver the right content at the right time, and manage a caseload efficiently. Works at the intersection of clinical protocol and cultural context. Measures success in meaningful member interactions and health outcomes.
The design challenge was creating a system that served all three stakeholders simultaneously, where the content felt personal to the member, the workflows scaled for the CHW, and the outcomes satisfied the payer.
Key Insight
This was service design, not screen design. The members we were trying to reach were precisely the people least likely to download a health app or visit a portal. The engagement channels had to meet people where they already were: their text messages, their phone, their email inbox, and their mailbox. The "product" was the orchestrated experience across channels, not any single screen.
What I Built
Research Systems for Cultural Sensitivity
Before we could build content or engagement flows, we needed to deeply understand the communities we were designing for. I built repeatable research systems that could be applied each time we entered a new population segment or clinical pathway.
The research focused on understanding cultural attitudes toward health, trust dynamics with healthcare institutions, communication preferences and literacy levels, family and community decision-making structures, and barriers to care that had nothing to do with the healthcare system itself (transportation, childcare, work schedule constraints, language, digital access).
This research wasn't a one-time project. I designed it as a repeatable capability so that every new clinical pathway or population segment would go through the same rigorous understanding process before any content was created.
The Annual Wellness Journey
The flagship product was the "Annual Wellness Journey," a year-long engagement program designed to guide members through preventive care milestones. Rather than sending isolated reminders ("you're due for a screening"), the journey created a narrative arc that connected health actions to the member's life, values, and cultural context.
The journey used AI to personalize content along multiple dimensions: language, cultural framing, reading level, channel preference, time of day, and motivational approach. A message about a diabetes screening might be framed around family responsibility for one member and personal independence for another, based on what the research told us about their cultural context and individual preferences.
Clinical Pathway Expansion
After validating the Annual Wellness Journey, I led the expansion into specialized clinical pathways. Each required its own research cycle:
Diabetes care
Ongoing management, medication adherence, A1c monitoring, nutrition counseling. Fully launched.
Pre-natal and post-partum care
Pregnancy milestones, birth planning, post-partum mental health, newborn care. Fully launched.
Childhood immunizations
Parental education, scheduling, addressing vaccine hesitancy with culturally appropriate information. Rebranded and integrated into the Annual Wellness Journey as the pathway for minors, with the dedicated research and content design retained and implemented.
Care for older adults
Research and design completed. We deferred the full standalone pathway due to competing priorities, but implemented an MVP that allowed clients to layer specific HEDIS, Medicare, and Medicaid quality measures a la carte on top of the Annual Wellness Journey as a short-term bridge.
Weight management
Initial messaging incorporated into the Annual Wellness Journey. The full standalone pathway was deprioritized given the rapid market shift toward GLP-1-based interventions and the coaching/co-pilot tools emerging around that space.
Kidney care
Research initiated but not completed before my departure.
The 2.0 Redesign
The MVP/1.0 Annual Wellness Journey was well-informed by behavioral science and evidence-based clinical research, but it was not personalized. Every member received the same content. It worked well enough to validate the model, but we knew personalization was the key to scaling impact.
I led a full mixed-methods research cycle to inform the 2.0 redesign: dozens of qualitative user interviews and hundreds of quantitative surveys. In the qualitative phase, we asked members to react to the current topics in the journey. Which provided useful information and which did not. Which felt credible. What they wanted more of, less of, and which topics we hadn't addressed that they were interested in.
Unexpected Finding
The qualitative research surfaced unexpected trust gaps. Members told us things like, "Why is my health plan asking me this instead of my doctor? That feels strange." That finding made us redesign the framing of every message to better explain why they were receiving outreach, why now, and why from their health insurance provider rather than a clinician. We also drew clearer lines distinguishing what should definitely be addressed with a provider.
The quantitative research drove the personalization model. We presented different message variations and key calls to action and asked members which they preferred (and why) and which were most likely to prompt them to take the desired action (click a link, take a survey, schedule an appointment). We collected member attributes up front (age, sex assigned at birth, gender identity, race, ethnicity, sexual orientation, geographic location, education, household income, primary and secondary languages, US citizenship status, and others) and segmented responses by those criteria to identify patterns.
Sometimes there was a clear winner regardless of any cultural or social criteria. Sometimes there were no discernible preferences. But sometimes patterns emerged along specific dimensions:
Generational patterns
Different age groups showed clear preferences for communication modality and had different baseline levels of trust in the healthcare system.
Cultural patterns
Members (especially women) who identified as Hispanic were more likely to prefer messages emphasizing community ("you are not alone," "ask your family and friends for help," "it's important to have people to talk to and lean on"). Non-Hispanic women in lower economic strata or urban areas were more likely to be young single mothers without a strong family or church support system and were more inclined to seek help from federal, state, and local programs or charities before leaning on friends or family.
Gender patterns
Men showed stronger preferences for messages framing preventive care around being strong for their kids or being able to serve as a caretaker for their loved ones.
We spent significant time giving our LLM models custom training and guardrails to ensure that personalized messages were sensitive and collaborative without leaning into stereotypes or biases. A preference pattern identified across a demographic segment had to be applied thoughtfully, not as a rule that assumed every individual in that segment would respond the same way.
Pathway Orchestration and Householding
The 2.0 redesign also introduced a sophisticated content orchestration layer. As we layered multiple clinical pathways on top of the Annual Wellness Journey, we needed global constraints on message volume and frequency, and a prioritization methodology for removing redundant or conflicting messages. A member on both the diabetes pathway and the annual wellness journey shouldn't receive two nearly identical reminders about an A1c check in the same week.
We extended the Annual Wellness Journey to include minors, primarily messaging parents about their children (except in cases where state law permitted direct communication with older minors on specific topics). We also introduced "householding," a feature that recognized when multiple health plan members shared the same phone number and consolidated messaging accordingly.
Householding in Practice
Under the old model, a parent with twins would have received two identical messages saying "your child might be due for a checkup." The new journey would send a single message: "Hey [Mom's name], [Kid #1] and [Kid #2] may be due for a 12-month checkup. Here's your pediatrician's phone number. Can we help you get that scheduled?"
AI Integration
I directed the integration of AI across three layers of the platform, working closely with our data science team from design through production monitoring. Our guiding principles: understand the business problem and the data before reaching for a model, understand the limitations and biases of any approach, build in time for engineering and data science to monitor and retrain models in production, and establish clear guardrails without smothering projects with governance.
NLP-Powered Auto-Response
Care coordinators were spending significant time responding to common inbound SMS messages (appointment scheduling, medication refill requests, benefit verification). We built an NLP system to categorize inbound messages and automate relevant responses, trained on thousands of real responses created by human agents during the first year of the platform's operation.
We used Hugging Face's open-source transformer models as a pre-trained base to convert messages into vector embeddings, then applied cosine similarity in high-dimensional space to find and group similar sentences and recommend the best response from a list of approved templates in our content library. Every recommendation was reviewed by a human (binary agree/disagree) during the validation phase. Once the system could reliably interpret inbound messages and apply the correct tags, it used a JSON lookup table to find the appropriate response and instruct our execution engine to send it. The system could also detect nonsense messages and created cases to route to human agents when it was unable to respond automatically.
Before: basic rules engine
A handful of automated responses for the most common patterns. The vast majority of inbound messages were routed to human agents for manual response.
After: NLP-powered categorization
Massively expanded the scope of automated responses and cut the number of cases routed to humans by over 90%. Retrained roughly once a month to add new response templates, with no significant model drift after launch.
Production monitoring
Built a monitoring dashboard in Snowflake for ProdOps to research questions, monitor model efficacy over time, and review uncategorized responses to identify new emerging patterns.
A Design Decision That Mattered
I made the call to limit automation to clearly transactional interactions and always provide a path to a human. We couldn't let automation undermine the trust and personal connection that made the entire engagement model work. Members should never feel they were being managed by a bot when they needed a person.
Generative AI Content Personalization ("Hoppy")
Within months of ChatGPT's initial release, I led the design of a generative AI tool (internally called "Hoppy") that used prompt engineering on the ChatGPT API (starting on 3.5, later upgraded to 4) to personalize member-facing content at scale. We built a user interface in Streamlit on top of the API that our content team used daily.
Inputs
SMS and email content for preventive health, chronic condition management, and health-related social needs from our content writers. Company and client-specific style guides for tone and voice. Market and UX research findings and jobs-to-be-done frameworks. Third-party clinical sources (e.g., American Diabetes Association guidelines, evidence-based guidance on preventive screenings).
Outputs
Personalized SMS and email messages tailored to individual members across age, gender, race/ethnicity, and other cultural factors, targeted specifically at encouraging members to take the intended call to action.
Bias mitigation
Some biases were present in the initial model output. A significant portion of our customization work involved teaching the model what was desirable and appropriate to personalize versus where it crossed into offensive stereotypes or assumptions. This was prompt engineering, not fine-tuning: we iterated on system prompts, guardrails, and review processes until the output consistently met our standards for cultural sensitivity.
Results
Platform and Operational Impact
Scale
- 2.8 million+ members engaged across 15 states in 25+ languages by the time of the GroundGame Health merger
- 3 clinical pathways launched beyond the flagship wellness journey, with MVP bridge for older adult care and active research on kidney care
- Redesigned the Annual Wellness Journey from 1.0 to a fully personalized 2.0 with intersectional content personalization
Efficiency
- Revenue doubled without growing headcount. Per-member economics improved dramatically as AI and automation handled routine work
- $200K annual savings from NLP auto-response, eliminating the equivalent of 4 full-time care coordinators' manual work
- 78.6% increase in community health worker productivity
Quality Measures and Cost Impact
The culturally-tailored engagement model drove measurable clinical outcomes for payer clients:
Quality Improvements
- 20% lift in select HEDIS quality scores for Medicaid populations
- 9% lift in HEDIS quality scores for Commercial populations
- 60% increase in Annual Wellness Visits for Medicare, with gains in other STAR ratings
- 60% AWV completion rate among 38,000+ members who hadn't seen a doctor in over 16 months
Cost Reduction
- 5:1 ROI demonstrated for a major health plan partner
- 27% reduction in inpatient admissions
- 18% reduction in preventable ER visits
Post-Merger Scale (GroundGame Health, 2024-2025)
In 2024, SameSky merged with GroundGame Health, combining the culturally-tailored engagement platform I had built with GroundGame's community-based organization (CBO) network and care coordination infrastructure. The combined entity continued to scale the model:
4.8/5
Member experience
32%
Medicare retention lift
374K+
Social needs solved
$43M+
Flowed to CBOs
What I Would Do Differently
The biggest lesson from building at this scale was about content operations. Culturally-sensitive content is expensive to create, validate, and maintain. Every new language, every new cultural segment, every clinical pathway multiplied the content matrix. If I were starting over, I would invest earlier in a content management architecture that made creating, reviewing, and updating content variants more systematic. We built that capability over time, but earlier investment would have accelerated pathway expansion.
I would also formalize the feedback loop from CHWs to the product team earlier. Community health workers were our most valuable source of insight about what was working and what wasn't, because they heard directly from members. We incorporated their feedback, but I would build structured mechanisms (weekly synthesis sessions, standardized observation reports) from day one rather than relying on informal channels.
Finally, I would push harder on establishing data partnerships with payer clients earlier. We eventually demonstrated strong clinical outcomes (HEDIS lifts, ER reductions, AWV completion), but getting access to the claims data needed to prove those outcomes required partnership agreements that took time to negotiate. Starting those conversations at the beginning of each client relationship, rather than after the engagement model was already proving itself through activity metrics, would have given us a stronger outcomes story sooner and accelerated expansion.
This case study describes product strategy, research methodology, and system design. Member data, payer client identities, and proprietary engagement content have been excluded. Quality measures and cost impact data are publicly reported by GroundGame Health. The engagement architecture, research frameworks, and AI integration approach are available for discussion in interviews.