One Designer + AI. Hundreds of Deliverables. — Vincent Wendy, AI Engineer

The conference grew while its design capacity stayed fixed. When Vincent Wendy, AI Engineer’s senior creative designer, prepared this talk, the team expected 6,000 attendees. By the recording, that number had reached 7,000. More than 140 sponsors, more than 300 speakers, and more than 600 sessions needed assets from one designer, working within a company of roughly twelve to fifteen people. 1:52

The deliverables ranged from stickers and track mascots to landing pages, speaker announcements, wayfinding, and digital signs. Each carried a different failure risk. A missing sponsor logo could leave a sponsor unrepresented; an incorrect schedule could misdirect attendees. Hundreds of graphics meant hundreds of opportunities for a small detail to become an event problem.

Wendy’s expanded design team consisted of himself, Devin, GPT, and Figma. His enthusiasm came from having concrete work for those tools: recurring assets, awkward handoffs, and mistakes worth preventing. The conference’s problems supplied the direction for automation.

A pelican riding a bicycle supplies the first example. Wendy recalls Simon Willison using that prompt to test models’ ability to produce a vector file. In Wendy’s own attempt, the basic model’s vector output was unsuitable for his design work. The practical requirement was a graphic he could use, so he proposed a different route: ask ChatGPT for a still image, such as a PNG, then vectorize it in Figma. 4:02

That changes which tool handles each step. Image generation supplies the picture; Figma handles the conversion into a vector asset. The model no longer has to produce the final file in one attempt. Choosing an intermediate format gives the designer a way around an output limitation while keeping the intended deliverable in view.

The method has five parts: establish the foundation, make designs reusable, automate workflows, validate output, and remove friction. The foundation is an ordinary design system of typography, colors, and components. Drawing on his product-design background, Wendy uses atomic design: define small pieces, then assemble them into larger deliverables like Lego. 4:52

That foundation supports two kinds of reuse:

• Rules for the agent: Defined desktop and mobile typography give Devin concrete choices to follow. Without them, a language model can invent font sizes and produce inconsistent work. Colors and typography settle decisions before generation begins.
• A reference for colleagues: Once the website expresses the branding, marketing can use it to create emails, flyers, and documents. The designer’s decisions become available through a product the team can inspect.

The mascots follow the same logic. A common treatment lets the designer repeat an established design rather than reconsider every asset from scratch. The initial work becomes useful across many outputs. Automation can then attach those layouts to the conference information that changes.

Room schedules used to be laid out manually in Figma. The replacement workflow pulls the latest data, selects the requested room and day, and exports the schedule as a PNG. Wendy downloads the image, carries it on a flash drive, and puts it on the room’s screen. Preparation becomes repeatable, while delivery still requires moving the image onto a physical display. 8:04

The tool separates changing schedule content from the layout used to display it. Another room or day can follow the same export path without another manual arrangement of sessions. Pulling fresh information also avoids relying on a previously composed graphic as the source of schedule content.

Visual fidelity still requires instructions and feedback. Earlier designer-to-engineer handoffs involved repeated corrections when an implementation missed the intended layout. Devin lets Wendy request a closer match and provide design information through MCP. If that route does not work, a specification sheet states spacing and font sizes directly. The shorter handoff helps, but the measurements supply what the agent needs to correct.

More than 300 speakers make individual announcement production impractical for one designer. The speaker-announcement tool lets a user select a speaker and change the displayed name. It supports a landscape layout, and a speaker’s headshot and details support automatic export when they are available. Trading cards extend the approach to different themes, including a treatment inspired by TBPN. 9:17

The headshot, name, and details vary between outputs; the intended spacing, type, and colors stay defined. Wendy describes the results as pixel perfect and explains the inputs that make this possible. His working loop runs through Slack and Figma, with Devin living in Slack. Design information reaches the agent through MCP or a specification document.

A free Figma plugin produces an annotated PDF with spacing, font sizes, and colors. Even frames with unhelpful names can contain useful measurements. The annotation carries essential visual decisions into the implementation request, rather than leaving the agent to infer them from a vague description. 10:49

For this production task, Wendy prefers the direct Slack–Figma–Slack loop to the longer design-thinking process he used earlier in his career. That judgment concerns his particular work: turning an established design into many faithful assets. The fast loop rests on earlier decisions about the foundation and on specifications that make those decisions usable.

Thumbnail creation exposes another bottleneck: finding a speaker in a photographer’s collection. Previously, Wendy inspected photographs one by one, sometimes using capture times to narrow the search. A new matching interface, which he likens to Tinder, helps check whether photographs depict the same person. In the example, the match is Jason Liu; Wendy can download that photograph and place it in a thumbnail. He considers the match accurate, although the example does not establish reliability across the collection. 11:47

The useful change is where manual searching ends. Associating the photograph with the speaker lets the designer move into thumbnail composition. Wendy’s excitement comes from the accumulation of removed chores: faster creation and less repetitive searching leave him looking for the next real problem worth solving.

Vision also helps inspect completed assets. The lobby banner contains more than 140 sponsor logos. Wendy asks Devin to check whether any logos are missing and reports 100 percent accuracy in his tests. That figure applies to those checks, whose size and conditions are unspecified; it does not establish a general guarantee of logo-detection accuracy. He uses the same approach on the conference T-shirt. 13:17

These applications give visual AI two distinct jobs:

• Asset retrieval: Find a photograph associated with the speaker who needs a thumbnail.
• Omission checking: Inspect a dense graphic for a required logo that a human might overlook.

A sponsor graphic can look finished while still missing something consequential. Combining human review with an AI check gives the designer another chance to catch that mistake.

Removing friction also means following the attendee’s journey. Wendy imagines arriving at registration, encountering wayfinding, using a QR code, and finding a room. Those pieces need to connect. A sign’s purpose becomes clearer alongside what the attendee just did and what they need to do next. The deliverables collectively help someone move through the conference. 14:15

The room-schedule workflow returns with a last-minute exception. One morning, a schedule needed updating, but the tool had no Edit button. It could generate an asset without providing the correction control Wendy now needed. He asked Devin to add the button, and it did. Wendy could change the schedule, export another PNG, and put the updated output back on the screen. 15:15

Where does a late correction enter the signage pipeline? The diagram shows the original export path and the added editing step. Fresh data supplies the room-and-day schedule; the new control lets Wendy change it before exporting again.

The change is small—an Edit button appears—but it makes the whole workflow more useful under event conditions. A correction can reuse the existing export and delivery steps. The PNG does not update the physical screen by itself: Wendy still has to deliver the new image. The tool handles repeated production, while the designer recognizes the exception and asks for the missing capability.

Wendy closes by recommending small-scale thinking for a large-scale problem: identify the details that can go wrong and address them early. His confidence in automation is expansive, but the examples make the advice concrete. Define typography, reuse a layout, locate a photograph, check for missing logos, and make corrections possible. A real problem becomes an advantage because it tells the designer which capability will help ship a better product. 15:46