I work upstream of the build: defining the core user problem, integrating AI into the workflow with structure and QA, and shaping how the result is positioned. Two live products and a portfolio of analytics work demonstrate how I think, not just what I deliver.
MS Marketing, Johns Hopkins Carey · ex-McCann Worldgroup planning · Washington, DC
One is a two-sided marketplace solving a concrete, under-served pain point. The other is a system for making AI dependable enough for serious work. Each is broken down the way I actually work: context, problem, insight, strategy, execution, QA, impact.
A verified, student-only subletting marketplace that closes the gap between 12-month leases and a 9-month academic year.
Trust-first homepage positioning: verified student housing, not another open marketplace.
Leases run 12 months; the academic year runs about 9. Every spring, students either eat the cost of summers they're not living there, or scramble to sublet through chaotic group chats where listings are inconsistent and scams are common.
The workaround — informal subletting — is its own problem. Listings have no standard, counterparties are unknown, and a single bad actor can cost a student their deposit. The friction isn't finding a place; it's trusting the person on the other side fast enough to act.
Sublets already exist; what's missing is a verified, legible channel. Students move quickly when they trust the counterparty. So the right move isn't another open marketplace to scroll — it's to standardize and verify, and make trust the product.
I scoped the MVP around a single job: get strangers to trust each other quickly. Everything else — listings, search, messaging — serves that. The positioning ("verified, student-only") wasn't a guess; it came out of a structured, AI-assisted research process with human judgment on the final call.
The live product keeps the network to verified students, standardizes how every sublet is listed, and cuts time-to-match. It's now live at subletu.xyz.
Before committing, I ran a Stage-2 validation: qualitative research on feature trade-offs and messaging across four angles, then a five-dimension brand-similarity matrix to test competitive fit. The verified-student angle wasn't kept because it sounded good — it was the only concept that held up on both trust and student relevance at once.
The research surfaced white space no existing platform holds, and the product shipped against it — starting with the Baltimore–DC corridor where demand is concentrated and predictable.
An AI workflow and QA layer that makes AI dependable enough for serious, high-stakes work.
Quality gates turn AI work into a staged review process before an output ships.
The default way people use AI for real work is one model, one prompt, one pass — then they trust the output. For anything that matters, that workflow is fragile, and the race is all on the generation side.
Weak assumptions go unchallenged, the wrong model gets the wrong job, and nothing catches what's confidently wrong. The failure isn't that AI can't generate — it's that nobody is checking the output before it ships.
Serious thinking separates research, writing, critique, and review into stages, each with the right tool and a check on the last. AI should work the same way. The differentiator isn't generation — it's the QA layer.
I set the thesis — structure beats single-shot prompting — and designed FOVRA around it: routed stages with a built-in critique-and-QA gate, positioned deliberately away from generic "AI automation" language. Making QA the core feature is what turns it into a defensible category instead of another chat tool.
FOVRA breaks AI work into routed stages instead of one long chat, and helps users pick the right model per stage. It's live in free beta at fovra.app.
The core of the system is a set of gates — input, prompt, output — each with a falsifiable rubric. A failed critical criterion blocks the deliverable; a failed warning flags concerns. A seven-dimension rubric scores every pasted output and flags issues by severity, so review is structured, not vibes.
By owning the step that decides whether an output can be trusted, FOVRA sits in a different category from the generators — and turns AI from a fast first draft into something dependable enough to ship.
Graduate and undergraduate projects showing the method behind the recommendation: cleaning real data, modeling it, and turning the output into a client-ready decision. Grouped by the method or tool each one leans on.
Cleaning, modeling, and interpreting data — from social-listening exports to multi-factor risk models.
Method. Cleaned a raw Sprout Social export in R — missing-value checks, grepl() filtering, dummy coding, ifelse() sentiment encoding, log-transform — then regressed post engagement.
Strategy. Costco shoppers debated deals (65%); Walmart skewed to emotional complaints (43.7%) → a differentiated 4-P playbook per retailer.
Method. Ordinal logistic regression, forward stepwise selection (AIC), and a pruned decision tree to test air pollution beyond smoking.
Finding. Air pollution is a conditional, non-linear risk factor — amplifying severity in high-obesity, high-exposure subgroups.
Decision tools that project trends, compare scenarios, and quantify trade-offs.
Method. Built an Excel decision tool projecting future mileage from a trend trajectory, comparing dedicated, one-way, and combined swapping scenarios.
Outcome. Quantified the combined program's savings and flagged where the one-for-one assumption limits the model.
Designing studies, prepping the data, and running the models that turn responses into pricing and retention calls.
Method. Designed a survey, dummy-coded responses in Excel, and ran regression in JASP to test price sensitivity, social influence, and theater-vs-streaming (n=25).
Strategy. A-List members prefer theaters significantly (β ≈ −2.06, p<.05) → hold them on value and identity, not price cuts.
Method. Conjoint analysis to estimate attribute importance and willingness-to-pay across visitor segments.
Strategy. Express coverage is the dominant driver (42.8%) → hybrid pricing with targeted bundles for families, IP fans, and thrill seekers.
Articles and research summaries I've co-written outside the classroom. Each one links to the full piece.
Leaders from biotech, pharma, finance, academia, and policy compared the US and China as two different paths through biopharma's innovation gap. The summary covers regulatory pathways, pricing and reimbursement, and why the US market stays central for global biopharma.
“For the future biopharma industry, the winning equation is not only relevant to China biotech, but to global innovators: strong science, disciplined execution, cross-market trust, and a shared commitment to improving health.”From the roundtable summaryRead on CBA ↗
// product marketing · product strategy · AI workflow design
I'm a marketing and product strategist with an MS in Marketing from Johns Hopkins Carey and a planning background at McCann Worldgroup. What sets me apart is how I use AI: to think better, not just produce more. I use it to turn ambiguity into clear workflows, pressure-test ideas, and keep human judgment on the final decision.
Most products answer the wrong question well. I start with the user's real friction and the assumptions everyone skipped, then work out where AI fits, how to route the work, and how to position the result so its value is obvious to the person choosing it.
AI moves fast; the person keeps control of the final decision. I design for that, not around it.
Routed workflows over one-shot prompts — research, drafting, critique, and review as distinct stages.
The step that decides whether an output can be trusted is the one I make first-class.
I'm drawn to roles where framing the problem matters as much as executing the answer. If that's how your team works, I'd love to connect.