Azure
Video AI
Clear, not easier: teaching developers a video AI platform
With our researcher, I defined a "clear, not easier" direction for the Azure Video Analyzer portal from a nine-session study with five developers, which turned a stalled naming debate with the API owners into a concrete stakeholder choice.
My part
Designer: AVA portal and widgets; Video Indexer account flows
Duration
2019 – 2022
Crew
PM, API owners, engineering, research, content design
Ships on
Azure portal
Context & role
Azure Video Analyzer (AVA) was a hybrid edge and cloud platform for capturing, recording and analyzing camera video. It launched in public preview in May 2021 and was retired in November 2022. Its target user was a solution developer building apps for site operators, described in the team's persona work as "not a video or AI expert". Azure Video Indexer is a separate service that indexes stored media for insights; for about a year the two products shared the "Video Analyzer" name. My role: I was the product designer for the AVA cloud portal and widget experiences from late 2020 to 2022, and the Azure Media Services-side designer for Video Indexer account flows from 2019 to 2020. I designed the screens and prototype, our researcher led the research method, and PM and engineering owned the decisions.
Stakes
The portal wasn't where customers would run AVA at scale; it was where first-time developers would learn the concepts before moving to the API. If that first impression was confusing, the product would lose developers before they wrote a line of code. The widget plan made the same bet: AVA's opportunity was to speed up adoption by reducing the domain expertise a developer needed. On Video Indexer the stakes were support cost: "Video Indexer can't connect to your Azure Media Services account" was one of the top support issues I designed against.
Reframe
Asked: finish the portal flows. Reframed: fix the vocabulary.
The brief was a feature list: topology, live pipeline, batch topology and job screens for Ignite 2021. The portal was being designed after the API, so it inherited the API's object model (PipelineTopology, LivePipeline, PipelineJob), names that describe how the system is built rather than what a developer wants to do. By August 2021 I had shown the new design to two of three planned groups, and every group struggled with the terms. I reframed the work from finishing screens to fixing the vocabulary and the order of steps, because no layout could fix words people didn't understand.
Illustration: the API's JSON object model, versus a visible flow that keeps the API's words.
Evidence
A nine-session study with five developers
With our researcher I ran a two-part study between August and October 2021: five 60-minute discovery sessions, where participants completed tasks in my Figma prototype while thinking aloud, and four 30-minute validation sessions comparing design options. It was small and qualitative, so it showed where people broke, not how often. Everyone could click through the flow, but comprehension and satisfaction were low, and every participant failed the "create a job" task. A flow-chart topology editor, where developers drag cameras, AI processors and outputs onto a canvas instead of writing JSON, drew a positive response from everyone.
01
"Topology" reads as networking. Adding "template" or "blueprint" beside it helped people understand what they were building.
02
The order was backwards. AVA asked developers to design a topology before connecting anything; they expected to connect cameras first and add rules last.
03
Nobody wanted a pipeline to trim a clip. Everyone looked for a tool on the video player, and three of five said batch analysis of recordings fit their budgets better than live.
In their words
"Why can't it be simpler?"
Expert customer in the AVA study
Key decision
Keep the API's words, make them clear
There were three options: narrow the persona to video and AI experts, who could cope with the current terms; rename the UI away from the API, as content design suggested ("live feeds", "template"); or keep the API's nouns and add plain supporting words, guidance and a visible order of steps. The pushback came from the API owners: "topology" is a standard media-industry term, the object names had already shipped, and "pipeline" kept the objects reading as one system. I chose clarity over renaming because the portal's job was to teach developers the API they'd use at scale, accepting that first-time users would still have to learn the word "topology". We recommended plainer supporting words ("template" beside topology, "post-processing" beside batch), visually distinct topology and pipeline flows, ready-made batch topologies and clipping tools beside the player. The principle: abstract the hierarchy and order, not the terminology.
Second call
Design two front doors for Video Indexer, not one
At the 2020 Video Indexer kickoff the plan assumed a single entry flow. I pushed the team to design two entry states: new users who might expect a free trial, and existing users who only needed to sign in to an account already linked to Azure Media Services. Treating them as one flow would have sent existing customers through onboarding they didn't need and left new ones without an explanation of what Video Indexer does. I also proposed a Video Indexer page inside the Media Services account with a plain explanation and direct routes to the Video Indexer portal and API, designed against the "can't connect" support issue.
Illustration: three answers to the naming debate, and why "clear, not easier" won.
Craft
Guided empty states and honest dead ends
Because a topology does nothing on its own, the empty state on the live pipelines page explains why a live pipeline is needed and walks through the two steps in order, with the first marked done. The account overview split three intents: try a demo, start building with tutorials, or go to next steps. Design reviews caught real dead ends early: widgets wouldn't work for uploaded files, so the video page needed a different state for them, and there was no link from a video back to its camera, so we removed the camera name rather than show a broken one. The portal also went through a formal accessibility test pass before the next release.
Illustration: guided empty states and honest dead ends in the portal.
System
One widget model and one vocabulary across products
The portal was one surface of a larger model. I contributed to the AVA widget requirements with the Video Indexer team: a single player for live and recorded video, an insights timeline, search and an area-of-interest drawer, each embeddable in a customer's app, themable, accessible to WCAG 2.1 and reusable across the Video Indexer portal, AVA apps and the Azure portal. Another Azure team later imported the AVA widget as its video display and zone drawer for AI skills. The same thinking applied to deprecation: when Media Services retired two analysis presets, I proposed a rule rather than a one-off, a page banner for a large deprecation and an inline message for a small one, chosen by how many customers are affected.
Illustration: one widget model, reused across three products.
What shipped
What shipped
Azure Video Analyzer launched in public preview in May 2021 with the portal areas from my design review: account setup, videos with the player widget, access policies and edge modules, with topologies and live pipelines added by August. The study's recommendations (plainer supporting words, visually distinct topology and pipeline flows, ready-made batch templates and clipping beside the player) became planned work in January 2022 as a lighter topology workflow and portal content changes. The widget model of player, insights timeline, search and area-of-interest drawer was built to be embedded and reused, and another Azure team adopted it as its video display and zone drawer.
Outcome
What changed
AVA launched in public preview in May 2021 with the portal areas from my design review: account setup, videos and the player widget, access policies and edge modules, with topologies and live pipelines added by August. The study gave the team its first evidence on terminology, with method and sample size, and moved the naming debate from opinion to a stakeholder choice; the recommendations became planned work in January 2022. Microsoft retired AVA in November 2022, about 18 months after preview, before portal usage could be measured.
Reflection
Name things with the API owners, before the API ships
I raised the naming problem after the core API names had shipped, so the best the portal could do was explain them. Next time I'd bring content design into API design reviews, recruit more non-experts (most participants were experts), and get telemetry live before preview, so a retirement decision has usage data behind it. The principle I took away: on a developer platform, the vocabulary is the interface, and naming is an architecture decision designers need to be in the room for. When two products share a name, customers pay for the org chart.
