01

The 30-second read

What I decided, and what it changed
The reframe
I changed the brief from “make video searchable” to “make AI answers verifiable.”
27 interviews showed users found answers but didn’t trust them. Verification moved into the MVP, collaboration moved out.
The hardest call
I made the AI's limits visible, and held that line against the CEO and Growth.
The CEO wanted every model capability on screen. Growth wanted the shortest onboarding. I shipped the narrower promise and kept verification in the first run.
What outlived me
After I left, the team shipped Video Decks on my design system, with no designer.
One component library for four surfaces, built as infrastructure rather than as my working file.
Results
Company reported
7,000+
Users within four months of launch.
Observed in testing
70% faster
To find a moment, vs. users’ own workflow. n = 27
Verified after handoff
3 decisions
Still in the product after the team rebuilt the interface.
Not measured
0
Experiments support the thesis, but I shipped without instrumentation.
02

The full story

7 chapters · about 8 min read
Role & context

What is VeedoAI

VeedoAI is an AI video understanding platform for legal, education, and enterprise teams. As the sole designer on a four-person founding team, I designed four product surfaces and the component library behind them, zero to 7,000+ users in four months.

Role
Director of Product & Brand Design (IC, No Reports)
IC · FULL AUTHORITY
Timeline
Feb – May 2025
4 MO · RESEARCH TO LAUNCH
Team
CEO, CBO, Director of Growth & Marketing
SOLE DESIGNER · 4 PEOPLE
Scope
4 surfaces, 3 segments, research, system & brand
0 → 1 · END TO END
Context & role

In a $23.8B enterprise video market, design was the only wedge a four-person team had

Every incumbent solved one slice: Twelve Labs had search without collaboration, AnyClip monetization without summarization, Gemini tagging without multimodal interaction. We could not out-model any of them.

Table comparing four products: Twelve Labs, AnyClip, Gemini, and Veedo AI beta; listing what each solves and their limitations. Twelve Labs solves AI video search but lacks collaboration. AnyClip provides video monetization without summarization. Gemini offers auto-tagging but no multimodal interaction. VeedoAI beta delivers answers verifiable against their source.

Sources: market size from MarketsandMarkets, Enterprise Video Market report (2024). Competitor positioning based on public product pages, Feb 2025.

That changed what my job was. I wasn't designing screens for an AI product. I was designing the company's differentiation — which meant every decision below is also a decision about where a team with no resources chooses to compete.

I owned
• All product design — research through hi-fi and design QA
• Brand identity and design language
The research program: 27 professional interviews
• The component library and prototyping
I influenced
• 12-month roadmap sequencing
• Feature prioritization against model limits
• Positioning and the differentiation narrative
I didn't own
• Chukwuma Nwaugha, CEO — model capability vs. interface promises
• Kit Desai, CBO — pricing surface and packaging
• Shamal Badhe, Growth — onboarding funnel and activation
Reframing the problem

The brief was "make video searchable." That was the wrong problem

27 interviews — 9 educators, 7 legal teams, 11 creators — showed the brief was aimed one layer too shallow. Legal and training participants said the same thing in different words: one wrong summary and they would stop using the tool.

Users weren’t failing to find things. They were failing to trust what they found.

So I changed what we were building: not a search tool with good AI, but a comprehension tool that is honest about being probabilistic. Search accuracy wasn’t the product problem. Verifiability was.

Comparison table with two columns: 'What They Said Was Broken' in red and 'What They Actually Needed' in green. Issues listed: Too much video, too little insight; Fragmented tools and workflows; Manual processes slow decisions; Knowledge trapped in video silos. Corresponding needs: Get the answer and see the second it came from; Act on the answer where I found it, without switching tools; Trust a summary enough to decide on it; Share a claim that anyone can check against the source.

The synthesis that moved the brief. Read the right column: every stated need is about acting on what the video contains, not locating it. That asymmetry told me the problem was one layer deeper than search.

A comparison chart with two rows and three columns showing planned roadmap and after 27 interviews results divided into MVP and Later Phase. The planned roadmap shows Collaboration under MVP and Verification under Later Phase. After 27 interviews, Verification was pulled into MVP, highlighted with a blue box, and Collaboration was pushed to Later Phase, shown with a dotted outline.

The reframe cost the team features it had already planned: verification affordances went into the MVP, collaboration surfaces moved to a later phase.

The metric I proposed

Nobody could say what the problem cost or what success looked like, so I built both numbers

Across 27 interviews one figure kept recurring: 3–5 hours per video lost to search and distribution. Nobody had multiplied it out. I did. It gave the founders a number to price against and told me which segment to design for first.

A cost calculation model showing the formula to estimate lost hours and yearly cost per user: 3-5 hours lost per video (n=27) times 4 videos per month (assumed 1 per week) times 12 months times $29-$36 per hour (BLS median wage, May 2024) equals $4.2K-$8.6K per year, before anyone acts on it. A note indicates it sizes the wound, not the cure, with a small label stating 'Modeled, basis stated.'

A summary that saves you four hours and is wrong once costs more than the four hours it saved.

There was no metric for “the AI understood this,” so I proposed three

Dimension 1 · Time
Time to first answer
Not time-in-player. How long from opening a video to having the specific thing you came for. This is the number the 70% improvement refers to.
Dimension 2 · Trust
Verification rate
How often a user clicks from an AI claim through to its source moment. The one I argued mattered most — and the one nobody in the category was instrumenting.
Dimension 3 · Error
Correction rate
How often the AI is wrong in a way the user catches and fixes. Low correction with low verification isn't accuracy — it's users not checking.
A horizontal scale titled 'Trust should be proportional to stakes' showing verification rate of user clicks from AI claims to source. The scale ranges from Near 0 (users trust blindly) on the left to Near 1 (AI saves no time) on the right. Two labeled points are marked: 'Creator: lighter' on the left and 'Legal: heavier' on the right, with a 'Healthy band depends on the stakes of the job' above them. Below the scale, a note states verification varies with trust posture: heavier for legal, lighter for creators.

I wasn’t designing to make people trust the AI more. I was designing to make trust proportional to stakes.

Product & key decisions

Four surfaces, one loop: ask, answer, verify

Ask a question against footage. Get an answer. Verify it against the second it came from.

Ask → Answer → Verify

One interaction connecting search, AI summaries, source moments, and recommendations across four surfaces.

Dashboard showing six video project thumbnails with titles, views, time stamps, and editing options.
Surface 01 — Projects

Comprehension starts before you open anything

Every project carries its summary at the list level — the first place "extract, don't watch" shows up, and the reason this isn't a file browser.

Video editing interface showing a makeup tutorial video with AI-generated summary and social platform options.
Surface 02 — Edit

The summary and its source never separate

Recaps generate from the timeline, not a detached transcript — so verification stays one click away at the point where a wrong claim would get baked into a shared clip.

Dashboard showing video analytics with project stats, engagement, watch time, and AI-driven recommendations.
Surface 03 — Analytics

Where attention actually went

The evidence layer under Surface 04. Without it, Refine would be the AI having opinions — the thing the rest of the product argues against.

Dashboard interface for video and blog content optimization with tasks and mark-as-done buttons.
Surface 04 — Refine

Advice carries its evidence too

Every suggestion links back to the behaviour that produced it — the verification rule applied to recommendations. One principle, four surfaces.

The live prototype

Start at the summary. Click any source chip to jump to the moment it came from. That round trip is the product, everything else on this page is an argument for it.

Open full screen in Figma ↗

The wireframes that led to the architecture call

Same screen, same chrome — one structural element changed. The captions record what each version taught, not what it looks like.

Before

Rejected — transcript-first
Wireframe display of a video player interface with search, keywords, topics, OCR, transcript, and scenes panel.
What it taught me
Participants could find words but couldn't act on what they found. Completeness without comprehension — and no amount of polishing the transcript layout fixed that, because the layout wasn't the problem.

After

Chosen — summary-first
Black and white sketch of a laptop screen showing a video player, search, keywords, topics, and transcript sections.
What it taught me
Swapping the content area while keeping every other element pinned proved this was an architecture change, not a redesign. That single observation shaped the final UI — and was the foundation for the before/after below.

↓ These wireframes became Decision 01 below — where the same choice is shown in final production UI, and the cost of making it is named.

Three calls, the options I rejected, and what each one cost

Before: raw meeting transcript, unstructured wall of text
After: AI-generated summary with decisions and action items
Before — Transcript After — Summary
01 · Information architecture
Summaries first, transcripts one click deeper
Every competitor led with the transcript. I bet on comprehension speed over completeness and made the summary the front door.
What it cost
If summaries proved untrustworthy, the architecture would need rebuilding. I took that risk instead of a layout that did both badly.
02 · Core interaction
Conversation against footage, not a search bar
Keyword search returns timestamps, not answers. I argued to the CEO that conversation was the moat, not feature count.
What it cost
Collaboration features (shared workspaces, commenting) moved to a later phase. Depth instead of coverage.
03 · Failure design
Making the AI’s limits visible, not hiding them
Where stakes are legal, trust means verifiability. I bet that admitting fallibility earns more adoption than performing perfection.
What it cost
A heavier interface, and an argument I kept having. The minimal version tested better at first glance and worse on second use.

Progressive AI disclosure — a timing spec, not a principle

AI products fail in one of two directions: they hide what the model can do, or they dump every capability on a first-time user. I rewrote progressive disclosure as what the interface must have proven by which second, and held all four surfaces to it.

Why it mattered beyond the screens: it gave a four-person team a checkable rule for arguments that would otherwise have been taste. "Does this earn its place before the ten-second mark?" settled scope disputes faster than any mockup.

Journey & validation

Every stage had a failure. This is the one I removed from each

Diagram showing a user journey with five steps detailing what broke and what was changed at each stage.
The stage that decided retention
Processing. A user who leaves during the wait never reaches the part that works, which is why the loading sequence got designed before the analytics did.
The stage that had to be earned
Explore. This is where the first AI error lands, and where verification either exists or the user is gone.

How I validated it — and what I couldn't

I ran moderated sessions on UserTesting and Lookback with all three segments, iterating between rounds. The fixes that worked all did the same thing: they showed where an answer came from.

The change that mattered most came late: participants who skipped first-run setup were the ones who abandoned after their first AI error. Onboarding wasn't polish — it was where trust was established or lost, and I rebuilt it to teach verification rather than tour features.

The honest limit of this evidence
We had no experimentation infrastructure. Nothing here is an A/B test — the 70% figure is an observation from moderated sessions, not a controlled result, and it's labelled that way throughout.
Text outlining an experiment with primary, cells, guardrail, and null hypotheses on verification and trust.
Impact after handoff

You don't have to take my word for any of this

Three things I can point at rather than assert. All publicly checkable, none of them mine to edit.

The interaction model persisted
I argued for conversational query over a keyword search bar, using a legal participant's own phrasing. The company's public product listing now describes the feature with the example question "When was the contract mentioned?" — the same question shape, still the primary way the product is explained to buyers.
The system kept shipping
Video Decks launched publicly on 28 June 2025, after my engagement ended and with no designer on staff. It is described as auto-distilling transcripts and visuals into concise slide summaries — the summary-first architecture, applied to a surface I never designed.
The reframe held in the market
An independent buyer directory classifies the product’s main job as research insights within video, not video search. A third party reached the same framing with no input from me.

Built as a system, so one designer could keep pace with four people shipping

The four surfaces share one component library — variants, tokens, handoff-ready. That's the only reason a solo designer kept pace with a founding team's roadmap: a new surface cost composition, not redesign.

The system outlived my engagement. The team kept building on the library after I left, without another designer on staff — the strongest evidence I have that it was built as infrastructure rather than as my own working file.

Desktop
Designed desktop-first. The ask → answer → verify loop needs the answer and its source moment side by side, so every surface was specified at desktop width.
Mobile
The core loop needs width. Rather than ship a degraded version, I scoped mobile to a later phase and took the reach cost knowingly.
Degraded states
AI-is-wrong was designed for. AI-is-slow and processing-failed were not fully specified — I documented them as known debt.
Locale
Multi-language transcription roadmapped later. Text expansion and non-Latin scripts were known debt I documented rather than solved.
Founder disagreements

Three calls I made, and the pushback I took for them

On a four-person team there is no design org to escalate to. I made these calls, gave the founders the tradeoff, and owned what followed.

Chukwuma Nwaugha
I changed the promise from “everything the model can do” to “what it does reliably.”
He wanted every capability surfaced. On a legal summary, one miss doesn’t read as a weak feature; it reads as a product that lies. Cost: we demo smaller than competitors who demo everything.
Kit Desai
I drew the pricing line on capability, not volume, so the free tier proved the thesis.
A flat price either shut out creators or lost money on heavy users. Basic search stayed free; advanced AI and faster processing went paid. A metered free tier would have felt broken before it felt useful.
Shamal Badhe
I kept the verification step in onboarding that growth wanted cut.
Testing showed users who skipped setup churned on their first AI error. I traded signup rate for activation quality. Cost: friction at the top of the funnel that growth had good reason to dislike.
Retrospective

What didn't move, what did, and what changed beyond the product

What didn’t move: verification rate, the metric I argued mattered most, was never instrumented. Users trusted visible sources in testing. I can’t show it changed retention.

What moved: four months from first interview to launch. After I left, the founders reported 7,000+ sign-ups across five industries, also cited on VeedoAI’s public ToolMage listing.

Beyond the product: “out-design, not out-model” became how the founders described the company. The design strategy became the business strategy.

Webpage for VeedoAI on ToolMage, showcasing a detailed description of VeedoAI, an advanced AI platform that transforms video content into searchable and intelligent assets featuring AI-generated slide decks, interactive chat, and transcripts. The page includes a 5-star rating, social media links, and buttons to visit or update the tool. A screenshot on the right highlights VeedoAI's features like video decks, contextual search, AI chat, frame explain, transcription, and smart scenes. Below, an overview explains VeedoAI's capabilities and user base of over 7,000 users.

What I'd do differently

01
Instrument the thesis before shipping it
I'd define a behavioural trust metric — verification-click rate against retention — before launch, so the strategy is falsifiable rather than merely plausible.
02
Narrow to the highest-stakes segment first
One core loop across three segments worked, but cost sharpness in each. I'd start with legal, let its verification requirements set the bar, and expand outward — sharpening later is harder than widening later.
03
Pressure-test onboarding in round one
First-run friction surfaced in late-round testing, when changes were expensive. In any 0-to-1, onboarding validation belongs in the first round, not in launch polish.

Before you ask in the interview

What was your specific role on VeedoAI?

Sole product designer on a four-person founding team, for four months. Director title, but an IC role with no reports: I owned research, product design, brand and the design system.

How were these metrics measured?

The 7,000+ is company reported and listed publicly on ToolMage. I didn’t have analytics access, so I don’t claim it as mine. The 70% came from moderated sessions (n = 27), not an A/B test, and I say so on the page.

What was the hardest decision, and who disagreed?

Showing the AI’s limits instead of hiding them. The CEO wanted to show off every capability, and Growth wanted a shorter onboarding. I held my ground, and the cost was a product that demos smaller than it is.

How did you design around model limits you didn’t control?

I assumed the AI would sometimes be wrong. Every answer links to the exact moment it came from, so users can check it in one click, and legal teams get more checking than creators do.

What from this applies to other AI products?

People don’t distrust AI. They distrust AI they can’t check. If your users act on what the model tells them, show them where every answer came from.

In a portfolio review

This page is the argument. Here’s what I’d walk you through live.

Trading signup rate for activation: the onboarding call I made against Growth
The cost model’s inputs, and where they were weakest
The transcript-first prototype, and why no amount of polish could save it
Portrait of a middle-aged man with short gray hair wearing large black glasses and a blue checkered shirt.

I had the pleasure of working with Olha during our time together at VeedoAI, where she served as Director of Product and Brand Design. From day one, Olha brought a rare combination of creative vision, design excellence, and strategic thinking that had a transformative impact on our product and brand.

Kit Desai
Co-Founder & CBO
Open to new opportunities

Ready To Start?

Book a free 30-minute discovery call

Tell me what you're building.
I'll tell you how I can help and exactly what it will cost.

Currently taking new clients · Typical start: 1–2 weeks from contract

Turning family data into confident action
Dashboard Design
2019 - 2022