Owned end-to-end UX: research, information architecture, interaction design, and usability testing.Collaborated with engineering on transcript parsing, time-codes, and clipping. Drove a lean build: shipped a usable v1 in 8 weeks and iterated after launch.

I followed a Lean UX approach with weekly iteration cycles, starting with research and moving into design and testing.

‍

I began by speaking with eight Account Managers and shadowing four real prep sessions to see the work as it actually happens. From those observations I mapped the end-to-end prep workflow—finding the call, skimming the timeline, capturing quotes, writing a summary, and sharing it with the client. In parallel, I audited the current storage setup to understand how files were named, who could access them, how long they were retained, and how clips were shared. I also benchmarked purpose-built tools like Gong and Chorus to understand where the real value is and what the smallest viable surface could be for our first release. The key insight was straightforward: AMs didn’t need “AI magic.” They needed a faster path to a trustworthy quote and a clean way to paste proof into their client threads.

The product vision centered on four principles. First, find fast: search needed to be the front door, with keywords, speakers, and tags immediately accessible. Second, proof points over paragraphs: one-click clips and timestamped links are more useful than long summaries. Third, repeatable: consistent labels and entity patterns would make notes comparable across AMs and accounts. Finally, privacy by default: redactions and access scopes should be built into the flow, not bolted on later.

The information architecture reflected the way AMs think. At the global level, users can search by query, date, and speaker, and quickly filter by client, AM, tag, or call length. The call view brings together a player and a time-coded transcript, with inline entity badges for competitors, prices, features, pains, and other cues that matter in follow-ups. A persistent clip tray lets AMs collect moments as they listen. When they’re ready to share, they can generate clip links, export a structured summary, or push everything straight into a CRM or email template.

Under the hood, each transcript is broken into sentences, which are tokenized and then labeled as entities such as competitor, budget, feature, objection, and next step. Clips are a lightweight object that stores the start and end times, the selected transcript range, and any tags applied during review. Personally identifiable information like phone numbers and email addresses—is automatically detected and masked in both the UI and any exports.

The interaction model is designed for speed. Clicking a sentence plays from that point; holding shift and dragging sets the bounds of a clip. As AMs listen, they can tag what they hear using single-key shortcuts (for example, c for competitor, $ for budget, n for next step). A share composer assembles selected clips and bullet points into a clean handoff that can be exported to the CRM or sent via email. We tested the flow with six AMs and iterated on the details larger hit targets for text selection, clearer clip boundaries, and persistent tags across sessions were the changes that moved comprehension and throughput the most.

For engineering, I provided specifications for every state processing, redacted, partial transcriptand documented interaction tokens and accessibility behaviors including keyboard operation, focus management, and captions. We partnered closely on performance, lazy-rendering transcript chunks to keep the UI responsive and using optimistic updates so clip creation felt instant.

In pilot, routine follow-up prep time decreased by roughly 38–45% compared to baseline. Seventy percent of sessions began with search or a tag filter, indicating that the “find fast” principle matched user behavior. Sixty percent of delivered summaries included at least one time-coded clip, improving clarity with clients. Post-launch, AM satisfaction averaged 8.7/10 in our internal survey. Risk also decreased because redactions are on by default and access is logged for compliance. (If you’re still collecting data, keep the directional framing and note that results are from the pilot.)

This project reinforced that decision surfaces beat raw storage. The winning pattern wasn’t heavy AI; it was faster proof—time-coded clips, consistent tags, and a share flow that mirrors how AMs already communicate. The model is now reusable anywhere we capture audio, from onboarding calls to research interviews, giving teams the same speed and confidence gains without increasing complexity.

‍