Portfolio · Selected work
Four projects, one question: does this work for real people, and does it build trust?
This is a selection of recent work across conversational and agentic AI, end-to-end product design, and UX research. Each case shows how I work: evidence before opinions, design for the full range of people, and trust treated as a feature, not a slogan. Some client names are withheld out of confidentiality; the work is described exactly as it happened.
- Agentic & conversational AI
- UX Research
- Product design
- Accessibility (a11y)
Agentic AI · Fintech · WhatsApp
Designing an AI agent to talk about debt, without making a hard moment worse.
The problem
Debt collection is one of the hardest conversations a company can automate. People arrive stressed, sometimes ashamed, often distrustful. A financial services provider wanted a generative-AI agent on WhatsApp to handle collections and refinancing, and the risk was obvious: an agent that pressures, confuses, or leaks data doesn't just fail the user, it damages the brand and can break the law.
What I did
I led the proof of concept end to end at Delto, a conversational AI startup, and designed the agent's behavior as a product, not a script:
- An empathetic conversation design compliant with Law 2300, which regulates how debtors may be contacted. Tone, timing, and wording were designed for a person under stress, not for a collections KPI.
- Guardrails as a design layer: an anti-fraud module (people legitimately doubt that "your bank" is really writing to them on WhatsApp) and a FAQ module grounded in a knowledge base (RAG), so the agent answers from facts instead of improvising.
- Human-in-the-loop boundaries: explicit rules for what the agent may negotiate on its own and when it must hand off.
- Data protection by design: secure file transfer (MFT/FTP, RSA-2048) for the sensitive data behind every conversation.
Outcome
The PoC delivered 9 of its 10 milestones. Beyond the numbers, it proved the point I care about: an AI agent can handle a delicate financial conversation while staying legal, honest about being an AI, and respectful of the person on the other side.
Related work at Delto: on a production collections agent for a health insurer, I analyzed 51,108 conversations from 45,359 customers running at 67% self-resolution, found roughly 60% of the value untracked, and proposed changes projected to add 430 conversions per quarter. Design decisions deserve evidence after launch, not just before.
End-to-end product design · Data visualization · Accessibility
Energy Hub: from field interviews to a high-fidelity design, for the teams electrifying rural Latin America.
The problem
The IDB needed a geospatial platform on rural electrification in Latin America and the Caribbean: a place where governments, funders, and technical teams could see where energy access stands and decide where to act. The data existed. What was missing was a product that made it legible for very different audiences, from policy makers to engineers in the field.
What I did
I owned the design end to end:
- Research first: stakeholder interviews, stakeholder mapping, benchmarking of comparable platforms, and co-creation workshops with the people who would actually use the tool.
- From evidence to definition: user personas and user stories with acceptance criteria, so engineering could build against needs rather than assumptions.
- From definition to design: wireframes iterated into high-fidelity mockups, with a UI kit and technical documentation for developer handoff.
- Accessibility and data visualization as first-class constraints: maps and indicators designed to be readable by the full range of people, following WCAG.
Outcome
A complete, documented design ready for development: research trail, personas, user stories, high-fidelity screens, and a UI kit that lets the platform grow without losing coherence. The kind of handoff where engineering says "we know exactly what to build and why."
Human-in-the-loop AI · SaaS · Copilot design
A marketing copilot that suggests, and only acts when you say so.
The problem
A marketing product built on Azure OpenAI needed AI features that content teams would actually trust. The temptation with generative AI is full automation; the reality is that professionals don't hand their brand voice to a black box, and they shouldn't.
What I did
I defined the scope and the behavior of the AI features end to end, not just the delivery plan:
- An AI copilot for content teams that evaluated and scored marketing pieces and suggested concrete improvements. The key design decision: the system executed changes only after explicit user approval. Human-in-the-loop wasn't a checkbox, it was the interaction model.
- A conversational FAQ agent for end users, grounded in a knowledge base (RAG over an LLM), so answers stayed accurate instead of plausible.
- An analytics module for market results, because a copilot that can't show its impact doesn't earn a second sprint.
Outcome
AI features that shipped with a clear trust contract: the copilot proposes, the human decides, the answers are grounded, and the results are measurable. I also managed the development with an offshore vendor, translating design intent across time zones.
UX Research · Service design · International organizations
CooperaNet: researching how a static site becomes a living cooperation network.
The problem
CooperaNet is the OAS platform for technical cooperation across the Americas. It worked as a publishing channel: content went up, and not much came back. The goal of the engagement was to turn it into an active hub where member states actually find each other, share what works, and collaborate.
What I did
I led the user research that grounds the redesign: understanding who uses the platform today, who should but doesn't, and what "collaboration" concretely means for government teams with very different capacities across the region. Research in this context has an extra constraint I know well from a decade with international organizations: every insight has to hold up across countries, languages, and political sensitivities.
Outcome
Ongoing. The research is shaping the redesign from a static catalogue toward a collaboration hub, with decisions traceable to evidence rather than to the loudest opinion in the room.
Contact
Products people understand and trust. That's where my work lives.
Whether it's an AI agent, a research engagement, or an accessibility audit, let's start with a conversation. I work remotely with teams across LatAm, the United States and Europe.
Book 30 minutesEmail: hola@silviadarosa.com · LinkedIn