Customer segmentation · personalization analytics
A synthetic, end-to-end work sample: framing the segmentation question, finding real structure in the data, naming audiences a business partner can use, and designing the activation and measurement that prove the work changed an outcome.
0 real records · 0 PHI · 0 proprietary sources
A 2-D projection of six behavioral signals across a synthetic base. Hover an audience to isolate its cluster. The structure is generated, not collected.
42k
synthetic records
member-like rows created only for this work sample
6
usable segments
named by behavioral need, not by statistical cluster number
18
activation plays
message, channel, owner, and KPI for each audience
0
real customer data
no company data, no PHI, no proprietary sources
·Point of view
A segment is not a cluster. If it cannot change a decision, it is a slice.
Audiences are defined by need and the lever that moves them, not by demographics alone.
Measurement is designed before launch: every audience gets a KPI, an owner, and a holdout where feasible.
Trust is a constraint. Segments stay explainable, privacy-safe, and fair to the people inside them.
01Approach
The goal is not a beautiful cluster plot. It is a reliable customer read partners can use to personalize messaging, prioritize experience fixes, and tell whether the work changed an outcome.
Start with the decision to be improved: which audiences differ enough in needs, language, channel behavior, and service risk to justify different activation strategies?
Assemble stable, explainable features from engagement, service friction, affordability, care complexity, and sentiment signals. Exclude variables that would create non-actionable or ethically questionable segments.
Use clustering and dimensional checks as discovery tools, then pressure-test whether the output is stable, distinct, interpretable, and useful for business owners.
Replace cluster labels with audience narratives: what this group needs, what makes them different, what message they respond to, and what intervention is worth testing.
Define the audience owner, campaign or experience lever, measurement approach, holdout logic, and feedback loop before the segment is put into market.
Where the workflow lands
Stage 01
Identity resolution, consent, and the governed feature refresh the segmentation reads from.
Salesforce Data Cloud · Adobe Experience Platform · Segment · Hightouch
Stage 02
Where this site's thinking lives: framing, clustering, validation, and the narrative that turns clusters into audiences.
Hex · Sigma · Mode · Tableau · Databricks Genie · Snowflake Cortex, with Claude / ChatGPT for the heavy compute and drafting
Stage 03
Audiences synced with the KPI, owner, and holdout attached, so the loop can close.
Adobe Target · Optimizely · SFMC Personalization · Dynamic Yield
Platform names indicate the category of tool each step lands in; the operating model is platform-portable by design.
02Synthetic customer view
These fields are illustrative. In a real enterprise, I would work with data owners to confirm availability, governance, refresh cadence, bias risk, and whether each signal is appropriate for the decision being supported.
| Family | Example signals | Why it matters |
|---|---|---|
| Engagement behavior | digital usage, email interaction, channel preference, task completion | separates self-service audiences from members who need guided support |
| Service friction | repeat-contact proxy, unresolved issue proxy, support fallback behavior | identifies moments where proactive support may prevent frustration |
| Affordability signals | cost concern proxy, plan-shopping behavior, renewal sensitivity | supports plain-language benefit education and retention planning |
| Care complexity | claims complexity proxy, condition-navigation proxy, benefit-use breadth | helps route members toward navigational help before friction occurs |
| Relationship sentiment | survey tone proxy, satisfaction trend, promoter/detractor movement | connects audience strategy to experience outcomes, not just response rates |
03Audience explorer
Every audience carries one stable color through the whole story: the same hue you saw resolve in the projection above.
Selected audience
This audience usually does not need more outreach. They need clearer digital next steps, fewer handoffs, and precise reminders at the moments where a task could stall.
22%
of synthetic base
Business need
Make the desired action easy to complete without requiring a service interaction.
Risk if mishandled
Over-communicating to this audience can create noise and lower trust in useful reminders.
Opportunity
Reduce avoidable service contacts by improving digital task completion and targeted next-best-action prompts.
Audience signature
Measurement
Worked example for Digital next-best-step nudge, read as incremental lift on the primary KPI against a randomized holdout.
Incremental lift
+2.2pts
Relative
+14%
Primary KPI
Digital task completion
Persona
“Give me the clear next step and let me handle it.”
Comfortable using digital tools, often satisfied when the path is obvious, and quick to disengage when messages feel generic or require a phone call.
Defining signals
Message principles
Owner partnerships
Activation
Digital next-best-step nudge
Trigger
Incomplete online task or benefit lookup
Channel
App, portal, or email
Recommendation
Use concise task language with one clear action, not a broad education campaign.
Measurement
Task completion, repeat visit rate, repeat-contact rate, and downstream satisfaction trend
Preventive reminder path
Trigger
Upcoming preventive-care window
Channel
Email followed by portal reminder
Recommendation
Frame the reminder as a simple checklist and make the scheduling path visible.
Measurement
Click-through, appointment-start proxy, and no-call completion rate
Self-service recovery
Trigger
Failed digital task or abandoned session
Channel
Portal banner plus service fallback
Recommendation
Offer a fast assisted option only after digital recovery has been attempted.
Measurement
Recovered sessions, escalation rate, and average contacts per issue
04Owner lens
The named-consultant part of the job is translation: the same segmentation lands differently for each audience owner. Here is how this read would be brought into each room before a campaign or program decision locks.
Audience owner
Older, benefit-intensive audiences where care complexity and service friction concentrate, and where the enrollment window compresses every decision.
Where the read concentrates
What the read changes
Which members get navigation-first outreach versus digital-first, and which get proactive claim explanation before the enrollment window opens.
Example decision
Pre-enrollment outreach prioritization: sequence guided-support contact for high-complexity members instead of sending one generic campaign to the full book.
Retention through the enrollment window · repeat-contact rate
Audience owner
Employer-group members who interact episodically, compare value at renewal, and expect consumer-grade digital experiences.
Where the read concentrates
What the read changes
Renewal communication defaults to concrete cost-and-value framing for planners, while self-servers get fewer, sharper digital nudges instead of broad campaigns.
Example decision
Renewal-season strategy: split the generic renewal journey into a value brief for cost-sensitive members and a low-frequency digital path for self-servers.
Renewal retention · digital task completion
Audience owner
Care and clinical-experience programs where the question is who benefits from proactive guidance and who is ready for preventive engagement.
Where the read concentrates
What the read changes
Which members are routed to care navigation before friction occurs, and which get preventive prompts timed to benefit windows rather than blanket wellness sends.
Example decision
Navigation capacity allocation: point limited outreach capacity at high-complexity members flagged by the segmentation instead of spreading it evenly.
Preventive action taken · resolved-on-first-contact
05Personalization read
Naming audiences is half the role; the other half is reading how each responds to different message strategies and turning that into the personalization defaults the enterprise runs on.
| Audience | Cost & valueconcrete tradeoffs, budget framing | Digital-firstself-service, one clear action | Guided supportreassurance, human navigation | Preventiontimely, benefit-aware nudges |
|---|---|---|---|---|
| Digital Self-Servers22% of base | +2.1 | +7.8best | -1.4 | +3.2 |
| Guided Care Navigators18% of base | +1.8 | -2.6 | +8.4best | +2.9 |
| Cost-Sensitive Planners16% of base | +9.1best | +1.9 | +2.4 | +0.8 |
| Preventive Health Engagers19% of base | +1.2 | +3.4 | +1.6 | +8.8best |
| Low-Engagement At-Risk14% of base | +3.6 | -0.8 | +4.9best | +1.1 |
| Claims Recovery Support11% of base | +2.8 | +0.6 | +7.2best | +1.4 |
Illustrative lift in percentage points vs. a generic control message · synthetic figures · hover a row to isolate
The diagonal is the program: match the strategy to the audience need and most of the personalization value is captured with four message systems, not forty.
Negative cells matter as much as positive ones: digital-first framing actively underperforms for Guided Care Navigators, so suppression rules are part of the recommendation.
Low-Engagement At-Risk responds weakly to everything, which is itself a read: the next test is entry point and channel, not more message variants.
06Measurement design
Segmentation does not stop at audience names. Each audience needs a measurable intervention, a partner owner, and a loop that decides whether the segment logic should be refined.
Primary metric
engagement lift, satisfaction movement, reduced friction, or a retention proxy, chosen with the owner before launch
Test design
randomized holdout where feasible; a matched comparison where a holdout is not possible
Feedback loop
segment movement, campaign response, and service outcomes feed back into the next refresh of the segmentation
Governance
privacy-safe features, explainable logic, and use cases appropriate to the experience being personalized
Activation ledger
What got activated, how it performed, and what fed back into the next refresh. The paused entry is deliberate: an honest ledger retires plays as readily as it scales them.
Renewal value brief
ScaledCost-Sensitive Planners
+2.4 pts retention proxy vs. holdout; renewal-window call volume down
Fed back: Cost-and-value framing promoted to the default renewal strategy for this audience.
Digital next-best-step nudge
RefinedDigital Self-Servers
+6.9 pts task completion vs. holdout; unsubscribe uptick in the highest-frequency cell
Fed back: Play kept; frequency capped and the cadence rule written into the audience definition.
Complex-moment guide
ScaledGuided Care Navigators
Repeat contacts down 18% vs. holdout; satisfaction trend recovered within the quarter
Fed back: Service scripts aligned to the same plain-language sequencing used in the message.
Relevance-first reconnection
PausedLow-Engagement At-Risk
+1.1 pts reactivation, not distinguishable from zero at this sample size
Fed back: Audience definition revisited: boundary members reassigned at the next refresh; channel-discovery test queued.
07Standards & coaching
Carrying the team’s segmentation standard means holding this line without reporting authority, and teaching it. A segment passes all six tests below. A slice fails at least one, and gets named as a slice, which is often still useful, just not the same promise.
Distinct enough to change a decision
Stable enough to monitor over time
Explainable to business partners without model jargon
Actionable through an owned channel, experience, or intervention
Measurable with a clean KPI and, where possible, a holdout group
Appropriate for customer trust, privacy, and fairness
How I coach the read
Before any clustering, the analyst names the decision the partner is facing and what a useful answer would change. If no decision is named, the work does not start.
Would a different feature set or a different k change the recommendation? If the story only survives one specific cut of the data, it is a slice wearing a segment's name.
One paragraph, plain language, consumable by a leader in ninety seconds. If the read does not survive translation out of model jargon, the segmentation is not finished.
08Limits of inference
Methodology rigor includes documenting where the read stops. A recommendation that hides its limits is a liability handed to the partner who acts on it, so the limits ship with the work, in the brief itself.
Cluster membership describes how groups differ; it does not establish why, or what will move them. Response is established by testing: the holdout designs above, not the clustering itself.
Members near segment boundaries can move between refreshes. Long-lived journeys should key on the behavior, not a hard-coded segment ID captured once.
Product changes, seasonality, and market shifts erode any segmentation. Refresh cadence, quarterly in this design, is part of the methodology, not an afterthought.
This segmentation is built for communication, experience, and service decisions. It is not appropriate for pricing, underwriting, eligibility, or any use with disparate-impact risk.
Lift estimates on an 11%-share audience need wider intervals and longer windows. The ledger's 'paused' entry is what honest uncertainty looks like in practice.
Synthetic separations are sharper than real customer data. In production, the same brief reports stability diagnostics, silhouette-style checks, and honest overlap between audiences.
09The deliverable
Everything above exists so a leader can read this in ninety seconds and make a call. This is the artifact the role ships: the read, what changed, the recommendation, and its limits, written to be acted on, not admired.
Customer & Decision Intelligence · Segmentation
The read
The base resolves into six stable audiences, and the actionable movement this quarter is concentrated in two of them. Cost-Sensitive Planners (16%) are entering the renewal window with elevated retention risk and respond strongly to concrete cost-and-value framing (+9.1 pts vs. generic in test). Guided Care Navigators (18%) continue to drive a disproportionate share of repeat contacts; the navigation-first guide reduced repeat contacts 18% against a holdout and should become the default for complex-claim moments.
What changed since the last refresh
Recommendation
Measurement
Each recommendation ships with its owner, primary KPI, and a randomized holdout where feasible; readouts return to this table in the Q4 brief.
Limits of this read
Reads are descriptive plus tested response, not causal claims about untested strategies. Small-audience estimates carry wide intervals; scope excludes pricing, underwriting, and eligibility decisions.
10Role alignment
My background combines quantitative data science, customer and campaign analytics, and hands-on AI-enabled building, grounded in degrees in mathematics, linguistics, and anthropology. That pairing is the job: the quantitative rigor the segmentation runs on, and the study of how people actually talk about their needs, which is what personas and briefs are made of.
Candidate context
8+ years quantitative / data science experience, including 4+ years in health-insurance marketing analytics: holdout-controlled campaign measurement, decile lift evaluation, enrollment and retention analysis, lead attribution, and causal inference. Segmentation-adjacent work that closed the loop from audience to activation to measured outcome. BS Mathematics; BA Linguistics & Anthropology. The persona and narrative craft in this sample is trained, not improvised.
This site was designed and built end-to-end in an AI-augmented workflow, the same class of toolkit the role leans on for the heavy compute.
01
Frame the question, design the approach, run it on modern platforms, interpret the result, and ship a recommendation audience owners can act on.
02
Read how audiences respond to different messages, offers, and experiences, and translate that into the personalization the enterprise relies on.
03
Act as the named segmentation consultant to audience owners and marketing leadership before a campaign or program decision locks.
04
Use modern AI-assisted analytical tools for the heavy compute while keeping methodology, privacy, and human judgment at the center.