iFind: AI-Assisted Video Editing for Behavioral Research
Running usability testing, UX research operations, and cross-functional feedback loops for a developmental research tool.
Project Overview
- Role
- UX Research Intern
- Organization
- Phil Fisher Lab (FIND Program)
- Timeline
- Nov 2023 - June 2024
- Methodologies
- Moderated Usability Testing (n = 50) · Think-Aloud Protocol · Heuristic Evaluations · Interactive Prototyping
Overview
The Filming Interactions to Nurture Development (FIND) program is an early childhood intervention program that records 60 minutes of caregiver-child interactions to coach caregivers on key developmental techniques defined by the lab (such as "serve and return" and "naming").
To scale this research, the lab developed iFind, an automated AI software designed to take the 60-minute interaction videos, use AI to identify developmental timestamps, and trim clips into program-defined structured coaching videos.
As a UX Research Intern, I ran usability studies to evaluate how participants navigated the beta tool, identified friction points in the AI-assisted editing flow, and optimized our team's research operations.
Key Contributions & Impact
Research Execution
Co-conducted 30 moderated usability sessions using the think-aloud method to evaluate beta software experience and intuition.
Insight Communication Efficiency
Built live Figma prototypes alongside qualitative notes, cutting build iteration turnover cycles from 12 users down to 4.
Educational Scaffolding
Designed visual reference graphics and testing cheat sheets for participants that resolved confusion around behavioral definitions ("naming" / "serve and return").
Research Operations
Standardized note-taking, built the team research repository, and created a shared scheduling system to eliminate testing overlaps.
01 — Context & Problem
The Challenge: Translating Domain Expertise to AI Workflows
Video editing in FIND's context requires strict adherence to lab coding definitions and coaching video structure. Early usability testing revealed two primary challenges:
Information & Domain Gaps
Participants struggled to grasp specific behavioral definitions (ex. "serve and return") during tasks, leading to editing errors unrelated to the UI itself.
AI Trust Gap
Users were hesitant to trust provided AI clip suggestions. Rather than wanting an automated tool that replaced their decisions, users wanted full agency to review, override, and manually fine-tune edits.
Before · manual editing
With iFind · AI-assisted
02 — Research Ops
Usability Testing Setup & Methodology
-
01
Sample & Recruitment
Screened participants across university mailing lists, tracking prior video editing experience to observe behavioral differences between novice and experienced editors.
-
02
Testing Protocol
Conducted 30 moderated sessions using a structured task script and the think-aloud method to capture live mental models, feature discovery order, and information architecture clarity.
-
03
Onboarding & Script Auditing
Regularly audited testing scripts & introductory slide deck to enforce consistency and precision in lab terminology (ex. standardizing "returns" instead of "responds").
03 — Key Insights & Changes
Users lacked confidence in AI automation without control.
Even though AI clip suggestions were not part of the tool components being tested, participants hesitated to rely on them and became unfocused, worrying more about the source and credibility of the clip suggestions and less about using them to complete the task.
The change
Easy editing, deletion, and addition of AI-identified naming instances; clip suggestions serve as a starting point rather than being the end-all be-all.
Conceptual confusion about FIND hindered task completion.
The first 6 participants struggled to differentiate "serve and return" in relation to naming during tasks despite verbal introductions.
The change
A provisional cheat sheet — goal definitions, clip structures, and behavioral examples. Subsequent participants completed tasks without hindrance of conceptual confusion.
iFIND's beta editing flow disrupted experienced users' mental models.
Experienced editors expected standard UI patterns — clicking timeline text bubbles to edit, download status indicators, default timeline templates, edited video in main player (not raw clips only).
The change
Fixed the heuristic violations (ex. missing download visibility), rebuilt the layout around pre-structured timeline templates and more intrusive top-anchored notifications, and created video playback toggle between original 60-minute session clip and edited video.
04 — Engineering Velocity
Unblocking the Design-to-Engineering Feedback Bottleneck
During initial testing, software updates lagged because engineers had to manually translate long document notes into visual UI changes. This meant that ~12 participants would test the exact same build before a fix was deployed.
My action — I used my design background to build a living, interactive Figma prototype that mapped exact interaction changes, animation styles, and layout shifts alongside our qualitative notes, removing stated interpretation guesswork pain point for rapid developers.
12 → 4
users per build iteration, after interactive prototype accompanied text-only notes
Reflection & Limitations
What I Learned
First, research ops is product strategy — standardizing the repository, scheduling, and note-taking formats is what keeps research quality steady across multiple facilitators. Second, bridging the dev gap with visuals for early software: handing engineers raw qualitative text created a translation bottleneck, and pairing insights with an interactive prototype is what actually accelerated implementation at the speed required of the lab.
Project Limitations & Future Opportunities
Unmeasured quantitative metrics
Early scope meant key measures like task completion and time-on-task weren't tracked systematically. Future rounds should add completion rates and SUS benchmarks.
Technological Advancements
With the state of AI at the time, high-fidelity prototypes were the extent of rapid visual comms I could produce. Today, I would leverage existing AI prototyping/dev tools to more accurately ship and communicate insights.