Inspiration

User testing today creates transcripts, not clarity.

Teams sit through recordings, extract notes manually, and still miss the most important signal: where real friction actually occurred.

We noticed three consistent problems:

1. Emotional frustration gets lost in text summaries

2. Visual context is separated from verbal feedback

3. Recommendations lack competitive grounding

MogUX was built to turn passive recordings into an autonomous qualitative analyst.

What it does

MogUX transforms user testing videos into an evidence-backed, self-improving UX playbook.

It operates through a three-layer intelligence loop:

  1. Detect Friction (Modulate)

We analyze the user’s voice for measurable emotional signals such as frustration, hesitation, and stress spikes. These serve as objective indicators that something in the UI failed.

  1. Correlate Context (Reka)

When a frustration spike is detected, MogUX analyzes screenshots from that moment to determine:

What the user was trying to do

What UI elements were visible

What likely caused confusion

This produces timestamped friction logs grounded in visual evidence.

  1. Benchmark & Recommend (Yutori)

For each friction event, MogUX autonomously researches:

 -> Industry best practices

 -> Competitor UX patterns

 -> Proven layout or interaction improvements

It then generates structured, benchmark-backed recommendations.

Over time, MogUX merges repeated friction points into an evolving UX playbook, increasing the priority of recurring issues and refining insights across sessions.

How we built it

MogUX is built as a modular agentic pipeline:

Modulate API → Detects emotional signals from audio

Reka Multimodal Model → Analyzes screenshots + transcript context

Yutori Research & Browsing APIs → Performs competitive benchmarking

LLM Orchestration Layer → Structures friction insights into actionable bullets

Delta Playbook Store → Performs incremental updates to maintain a living knowledge base

Architecture Flow

User Testing Video → Emotion Spike Detection (Modulate) → Visual Context Analysis (Reka) → Benchmark Research (Yutori) → Incremental Playbook Update

We intentionally structured the output as bullet-level friction logs instead of long summaries to preserve qualitative detail.

Challenges we ran into

Multimodal Alignment

Correlating emotional timestamps with screen state required designing time-window batching rather than relying on exact millisecond precision.

Avoiding Generic Summaries

Large language models tend to over-generalize UX feedback. We solved this by enforcing structured, localized friction bullets tied to evidence.

Scope Control

Automatically modifying UI code would introduce instability. We instead focused on high-confidence qualitative recommendations for human designers.

Designing Real “Self-Improvement”

We avoided fake learning claims by implementing incremental delta updates that merge recurring friction into a prioritized playbook.

Accomplishments that we're proud of

Successfully chained Modulate → Reka → Yutori into a coherent autonomous loop

Built multimodal friction detection grounded in emotional and visual signals

Created an evolving UX playbook that strengthens as more sessions are analyzed

Designed a structured output format that designers can immediately act on Most importantly, we demonstrated that user testing can move from passive observation to active intelligence.

What we learned

Emotional signals are a powerful and underutilized UX metric

Multimodal reasoning dramatically improves qualitative insight quality

Autonomous research agents increase credibility of recommendations

Self-improving systems do not require retraining — they require structured memory and incremental updates

What's next for MogUX

Real-time live session monitoring instead of post-session analysis

Cross-session clustering to detect systemic product weaknesses

Team dashboards with severity scoring and prioritization

Direct integration with product management tools (Linear, Jira)

A feedback loop where designers validate recommendations and refine the model’s prioritization

Our long-term vision is a world where user testing sessions automatically strengthen product intelligence without adding manual overhead.

Built With

Share this project:

Updates