RADARs

The user and stakes

“Be responsible” is not an actionable product requirement.

Responsible Innovation Labs asked our team to research and develop a tool for AI startup founders and investors to examine risk, strengthen trustworthiness, expand responsible adoption, and help shape regulatory conversations.

Early-stage teams face expectations around privacy, safety, fairness, transparency, and accountability, but much of the guidance assumes legal, policy, and governance resources they do not have. Founders needed a practical entry point: what matters to this product, why now, what to do next, and how to communicate current practice honestly.

The research recovery

We stopped asking founders to care more.

Our original research plan depended heavily on interviews with early-stage AI founders. After 400+ cold-outreach attempts yielded too little participation, I led the research-strategy recovery and we broadened discovery through investor, technical, design, healthcare, policy, founder, and Responsible AI perspectives.

The synthesis revealed a product constraint, not a values problem: founders could care about Responsible AI while survival, sales, and immediate product work still came first. I led the synthesis and pivot narrative that helped the team choose transparency as an actionable entry point.

That decision shifted the product from broad guidance toward guided decision support: tailored questions, explanations of why they mattered, visible input-to-output relationships, a human-editable report, and prioritized next steps.

Transparency became the connective layer between current product practice, customer trust, and the next decision a team could actually take.

What I actually did

My role and decision authority

I functioned as the strategy lead on a four-student team whose members all brought product or UX backgrounds. We worked collaboratively; I led the core framing and product-strategy workstreams while contributing across delivery and evaluation.

Trust, translated

The decisions that changed the product

From decision to system

A hosted prototype built to make the direction testable

Without a dedicated engineering team, we used a low-code stack. The workflow evolved across the capstone—from early RAG exploration, to a Webflow/Zapier/Sheets/OpenAI prototype plan, to a later Firestore, rules, model-assisted generation, CMS, and PDF handoff.

1. 01

   Onboard

   Capture domain, base model, core function, intended use, audience, and current Responsible AI practices.

2. 02

   Tailor

   Map product context to relevant questionnaire modules across privacy, fairness, safety, misuse, and synthetic content.

3. 03

   Assess and explain

   Combine answers, rule-based scoring, and model-assisted analysis to assemble risk areas and recommendations.

4. 04

   Review, edit, and export

   Render a draft disclosure report that the user can inspect, modify, and download before external use.

What the evidence says

Seven formative tests showed exactly where trust broke.

The topline numbers helped locate friction; the questions underneath them drove the product changes.

85%

reported task completion in the formative study

1.2

reported errors per participant

7 min

reported time per module

4.2 / 5

reported average satisfaction

• People struggled to see how early inputs mapped to questionnaire modules and report outputs.
• They questioned sensitive-data handling, the trustworthiness of generated content, and what scores meant.
• Guidance, progress, and download controls were too easy to miss.
• The team prioritized clearer progress, terminology, input-to-output mapping, report structure, next steps, and explanations for requested data.

Outcome and impact

A broad brief became a concrete, testable product direction.

The team delivered and handed off a hosted interactive application prototype plus technical reconstruction guidance. A real user could complete onboarding, answer tailored questions, wait for report generation, review and modify the draft, and download it.

The product and team impact is clear: a recovered research plan, a shared transparency-first direction, a testable workflow, and evidence-backed changes. The scoring and AI guidance remained partially simulated or unvalidated, and sponsor continuation after handoff is unknown.

Strategy impact

broad brief → transparency wedge

Product impact

hosted human-editable prototype

User evidence

seven-person formative study

Handoff

prototype + technical guidance

The next honest release

What I would build next

1. Introduce least-privilege access, explicit retention, and consent for sensitive inputs.
2. Add deterministic checks, citations, and source-to-recommendation mappings.
3. Calibrate confidence and make uncertainty visible throughout the report.
4. Require human approval before disclosure and add audit logs and regression tests.
5. Validate the assessment logic with qualified Responsible AI, legal, and technical reviewers.