NineFive was a B2B SaaS to help freelancers manage multiple clients in parallel — workspace, files, updates, feedback. Inside it, we shipped Wanda, an agent that answered client questions on freelance projects autonomously, with the project's full context as background. We were wrong about the market, and right about the agent.
Role
Founder & Product Designer
Products
Entrepreneurship - B2B SaaS
Dates
2022 - 2024
Building Wanda — a client-facing LLM agent in 2023, when "agent" wasn't yet a category.
Five clients, one freelancer, the same questions on repeat.
Freelancers managing 4-6 clients at once weren't drowning in work — they were drowning in the same questions, asked five times, by five different people. "Where are we on my project?", "Did you receive my brief?", "Can you re-send the latest mockup?". Each interruption pulled the freelancer out of deep work and back into context-switching mode.
NineFive's product hypothesis was simple: if we could absorb most client questions before they reached the freelancer, freelancers could manage twice as many clients without burning out. Wanda was that absorption layer.
Designing an agent before tooling caught up.
Wanda was an LLM agent built on GPT-3.5 via the OpenAI API, with a custom RAG pipeline that pulled context from each project's documents, files, and threads. Clients asked questions in their portal, Wanda answered them with the project as context, and the freelancer was only pulled in when the question required a human decision.
In 2023, this was harder than it sounds today. RAG tooling was nascent, vector databases were expensive, and token costs were a constant operational drag — we spent a meaningful share of engineering hours just trying to reduce the context passed per question without breaking the answer quality. The same problem a junior dev solves with a one-line LangChain call today took us weeks of architecture decisions then.
That constraint shaped the product. We designed Wanda to be economical by default: short responses, narrow context windows, fall-through to the human as soon as confidence dropped. The design wasn't just UX — it was also cost engineering. You can't separate the two when every token has a unit cost.
Live in production, on real freelance projects.
Wanda was live inside NineFive's client portal, answering questions from real users on real freelance projects, in production, for over a year. The interactive demo above walks through the workspace and Wanda's client-facing surface.
Interest, but not stickiness.
The launch metrics from NineFive in 2023 reveal the harsh reality of early-stage product fit. We hit an 18% visitor-to-signup rate — proof the value proposition resonated. But only 43% of signups activated, and despite a 54% retention rate among active users, our base plateaued at 23 weekly active users with volatile daily engagement.
We had proved interest, but not stickiness. With no compounding traction and a depleted treasury, we ceased operations.
What NineFive was actually preparing me for.
"Wanda wasn't a smart support bot. Wanda was an early version of what is now becoming the dominant pattern of 2026 software: an agent that orchestrates other agents inside a workspace dedicated to a customer."
We didn't have the language for it then. The tooling wasn't mature, the cost structure was punishing, and "AI agent" was barely a category. That's the lesson I carry into every product conversation today. Designing for AI agents isn't about adding a chatbot to a product. It's about reorganizing the product around the moment the customer interacts with the AI — what they ask, what the agent can actually do, where the human stays in the loop, and what the agent costs to operate. NineFive gave me three years of head start on a question that most product teams are only starting to ask now.