I was part of LinkedIn’s earliest explorations in generative AI, collaborating with a small, cross-functional team to identify ways LinkedIn could thread AI into existing consumer products. This work locked funding for further work and ultimately establishing a blueprint for how product teams build in these spaces.
Greenlit
Greenlit — The concept became funded product work
And informed future AI work at LinkedIn
LinkedIn Premium AI
LinkedIn Premium AI — Concept translated into consumer product
Both in 2023 and beyond
1 patent
1 patent — Published in 2026
Listed as an inventor on the patent
Hi Dana — I can help you find your next role and reach your career goals.
Here’s what I know about you so far.
This role centers on demand planning and vendor audits — owning the forecasting process end to end and simplifying reporting across accounts.
Based on your profile and preferences, here’s my assessment of the role at Meridian.
Agent identity and interactions across onboarding, recommendations, and threaded conversations
Author’s note: This case study covers a snapshot of the experimentation and exploration involved in this work. Screens and flows have been recreated to account for internal or unpublished features.
ChatGPT’s release in late 2022 had a profound impact on the tech industry, from product roadmaps to operating models and everything in between. I joined LinkedIn’s early exploratory work around generative AI in 2023 and had a direct role in defining how the technology could meaningfully thread throughout existing consumer products — focusing specifically on the job search experience. This work ultimately aligned executive priorities and funding, paving the way for future AI work from 2023 onward.
Generative AI was new territory for a lot of our team and our users. Along with that, the goal of this work was to create real, product-rooted concepts that worked with LinkedIn’s existing ecosystem. And it was all framed within a focused, 3-month timeline.
I collaborated directly with my design partner on this work — and in close partnership with product and engineering leadership. The two of us split the work across the platform and the conversational interactions, though the nature of the work meant most of the designs went through shared iterations. My focus was the conversational framework and the engineering partnership needed to productize it.
We’d established that the prototype would revolve around conversational AI affordances, with a real, personified agent working with the user through every step of the flow. My first goal when joining was to align the team on what that agent would be: its identity, its personification, and how it interacts with users across modalities.
After a few rounds of iteration, I aligned the team around a balanced approach: the agent would have a distinct identity and would collaborate in a warm, direct, yet measured way. It would be personified enough to support career work — but it wouldn’t lean into character or personality.
The prototype revolved around a series of core capabilities, and the conversation design — along with the traditional UI design — was built specifically for those conversation beats. And we opted for a “show, don’t tell” approach, backed by UXR.
Though I can’t share all of the specifics around its identity, the framework behind the agent’s interaction model is the same one that’s translated into products like Hiring Assistant and Premium AI features today. For example:
The proof of concept was the first real test of how this kind of work translates from an idea + intention into a real product with real contexts and constraints.
This is a snapshot from the onboarding flow. I proposed leveraging the moment to introduce the vibe and dynamics of what this agent-guided experience would feel like — and pairing the agent’s personal intro with a trust-focused confirmation.
Recommendations or insights from the agent were rooted specifically in the user’s personal context. It didn’t simply push recommendations; it framed them as qualified recommendations tailored to personal context.
While the concept’s scope covered the full job seeker experience, we also designed ways for users to dig deeper at each stage. For instance, digging deeper into recommendations and job fit, or learning more about a company’s values.
Throughout the experience, the agent tracked and recommended tasks related to the user’s job search — demonstrating one of the ways an agent could take on tasks while freeing up time for the most important parts of the search.
Hi Dana — I can help you find your next role and reach your career goals.
Here’s what I know about you so far.
This role centers on demand planning and vendor audits — owning the forecasting process end to end and simplifying reporting across accounts.
Overall, this looks like a strong fit for your skills and preferences.
Qualifications: Your four years in demand planning map directly onto the forecasting work this role owns. Your vendor-audit background covers the compliance half of the job.
Growth: Analysts at Meridian typically step into a lead role within two years, with a rotational program after the first.
You have 3 of the 4 core skills for this role.
On your profile
Missing
Common among other applicants
Forecasting tools · Advanced Excel · ERP systems
Prepare and apply
My focus was the architecture underneath each of these beats: what the agent says versus what the interface holds, what earns a full screen vs. a card vs. a single message, and what balance of information is most helpful at each step. The actual script within the chat experience was the architectural framing that tied each of these beats + questions together.
Alongside the design work, I also partnered with our engineering team on the technical side of this prototype: the backend structure that would enable the capabilities included across the experience.
This work specifically involved prompt engineering for the agent and the conversational framework — along with defining an engagement model for how we’d collaborate as a team longer term. I owned the identity and output elements of the meta prompt, and iterated through evals against the core requirements for successful outputs.
You are a career agent working within a professional app. You help users find roles, assess fit, and prepare for applications and interviews. You work with users based on their specific contexts and provide new ideas or directions when the situation warrants it.
You use the voice of a trusted career partner. You use a warm, direct, and helpful tone in your responses. You use plain language instead of jargon. Explain your reasoning in one breath, not three. Never perform enthusiasm; never scold.
Never claim to be a person, and never imply feelings you don’t have. Don’t speculate about a member’s prospects beyond what their profile supports. When you are uncertain, say so plainly — and say what would make you certain.
Lead with the answer. Roles, companies, and assessments return as structured entities, never prose. One question at a time. Every assessment ends with a next step the member can take.
The identity work, written as instructions the model could follow — not brand guidelines it couldn’t parse.
“Is this role worth applying to?”
Yes, and you’d likely be a good fit for it. You have 3/4 recommended skills, and the one you’re missing — Tableau — is similar to products you’ve used before. I can help you get started with a referral message or the application.
Absolutely! This is an incredible opportunity and you would be the perfect candidate. You should apply now — I think you’d love it.
| Evaluation criteria | A | B |
|---|---|---|
| Accuracy — every claim traces to the member’s profile | ✓ | ✕ |
| Identity — warm and direct; nothing performed | ✓ | ✕ |
| Boundaries — first person only where the agent did the work; no feelings it can’t have | ✓ | ✕ |
| Task focus — ends in a next step the member can take | ✓ | ✓ |
Every demo scenario carried a rubric like this. Responses were run, scored, and reported back — and the prompt guidance tightened on each pass. This work scaled across every capability in the prototype, and the evaluations helped identify what aspects scaled well and which ones needed task-specific tailoring.
The goal of this 3-month exploration was to create a proof-of-concept prototype that demonstrated the ways generative AI could thread into existing consumer products. The design work set out to demonstrate the vision, and the engineering work aimed to prove that the ideas were viable within the company’s tech stack.
The work did exactly that. Before the end of the exploration period, the prototype had aligned leadership and product teams on a genuine foundation for building AI products moving forward. The work we did — the interaction patterns, the prompt engineering, the conversation design — became the model for other teams to build on across the company. More importantly, the collaboration dynamics became a template for product teams moving forward.
Identity is architecture. It informs not just how a product sounds or feels — but how it functions. It sets the parameters and expectations for what an agent’s role is throughout an experience. And it needs to be durable and scalable at a prompt level.
Collaboration is invaluable. So much of the identity and conversation design work relied on the working relationships I set up with engineering and product. Learning from each other and working toward a shared benchmark helped prove that these ideas were viable and durable.
Direction over perfection. Given the context of this work, there wasn’t a realistic or attainable version of perfection for what we were building. Instead, the goal was vision and direction — things that could align the company on an idea and set a trajectory that people could understand and be a part of. Orienting around this regularly helped keep the work moving forward in a way that served us and the final prototype.